This PDF file contains the front matter associated with SPIE Proceedings Volume 11584, including the Title Page, Copyright Information, and Table of Contents.

High altitude balloon is a kind of aerostat working in the near space. Based on the basic radiation theory of physical optics, this paper takes the high altitude balloon as the research object. On the basis of summarizing and analyzing the research status of the high altitude balloon. Through the study of the thermodynamic model, the infrared radiation characteristics of the high altitude balloon are simulated. Through the study of the characteristics of the environmental factors of the high altitude balloon in the adjacent space, the infrared radiation of the environment of the high altitude balloon is simulated. Through the study of common atmospheric models, the atmospheric transmittance in the process of infrared radiation transmission is simulated. Finally, based on the linear quantization method, the radiation value is transformed into gray value, and then the infrared simulation image of high altitude balloon is obtained. The research and simulation results of this paper have certain reference value for the infrared detection and recognition of the high altitude balloon target.

Model iteration with new data is important to improve generalization of the model. In general, there are two methods to deal with model incremental update: (a) retraining the model with merging all data together and (b) training a separate model with the new data based on transfer learning. However, the above methods are either time-consuming or suffering from over-fitting problems when the sample size of new data is small. To address this practical issue, we propose a new iteration model, the IterationNet, which can learn features of new data while maintain the performance on the old data. It is a new model iteration method based on knowledge distillation which adds consistency network and truncate L1 regularization. In classifying fake avatar images of Weibo users, IterationNet extremely decreased training time from 8 hours to 5 minutes while the accuracy rate is only reduced from 96% to 91% comparing to training with merged data. Compared with transfer learning, IterationNet showed increased accuracy rate by 21 percent with similar training time.

Urine sediment detection is of great significance as one of the routine testing items. The traditional urine sediment detection method is mainly manual microscopic examination. Thus, it incurs heavy human workload and complicated operation, while it is easy to miss the targets. To alleviate this problem, a two-stage urine sediment detection method is proposed in this paper. More specifically, the segmentation and classification tasks are transformed into object detection tasks, and the feature extraction is performed by Deep Convolutional Neural Networks (DCNN). In our method, HOG+SVM is used as region proposal, and Trimmed MobileNets is used for DCNN refining. The experimental results demonstrate that the proposed method achieves promising performance.

This paper proposes a novel generation adversarial network (GAN) based on UNet++ architecture for infrared and visible image fusion. The idea of this method is to establish an adversarial game between the generator that generates the fusion image and the discriminator that determines whether the fusion image meets the standard. The generator uses the structure of UNet++, which does not have a deep network structure but establishes a dense connection at the shallow layer, so it has strong ability to obtain shallow features. As for the discriminator, it adopts a network structure similar to the Visual Geometry Group (VGG) network. The loss function uses the method of comparing the high-frequency part of the fused image with the two source images to make the fused image have more high-frequency information. In terms of high-frequency detail extraction of source images, two gradient extraction methods based on different combinations of directional extraction operators and high-frequency extraction are proposed in this paper. This paper compares the two methods above with other fusion networks, and proves that the fusion image generated by our network is highly similar to the infrared image, and has a lot of gradient information of the visible image.

Sea-Land segmentation based on surveillance images is an important research content for real-time coast monitoring. However, the complex weather and environmental makes the segmentation of sea-land is a difficult task. Although previous deep learning methods based on convolutional neural networks have achieved excellent results in semantic segmentation, and there has been some work using deep convolutional neural networks for Sea-Land segmentation but we hope that the image segmentation model can achieve more accurate results in sea and land segmentation. In our method, we propose a novel sea-land segmentation framework called Multi Sea-Land U-net (MSLUnet), the framework base on a multi-scale. The proposed MSLUnet is mainly composed of a multi-scale layer and U-Net convolutional network. The multi-scale input layer constructs an image pyramid to accept multiple levels of image data in the network model. U-shaped convolutional networks are used as the back-bone network structure to learn rich hierarchical representations. Experimental results show that compared with other architectures, MSLUnet has achieved good performance.

The quality of the images collected by the coastal zone video surveillance equipment is seriously degraded due to the sea fog, which directly affects the analysis of the image. Therefore, the study of the costal image dehazing method is of great significance to the related research of the coastal zone. Costal image has the characteristics of large sky area and monotonous color. The traditional method based on atmospheric scattering physics model is not suitable for this kind of image for block effect and color distortion. In this paper, we introduce the generative adversarial mechanism into sea fog image defogging, and propose a coastal image dehazing network based on it. The proposed model includes a generative network and a discriminative model, and is trained by adversarial mechanism. The generative model is composed of multi-scale feature extraction module and residual connection module. The discriminative network consists of two subnetworks of receptive field of different sizes.

Nowadays pointer meters are widely used in industries. In this paper, we propose a novel automatic reading method for pointer meters. Different from previous works that adopt two-stage detectors for detection, our approach integrates two components: a one-stage detector NeighborNet (Neighbor Network) and an angle reading module. NeighborNet is a simple but effective framework that can generate keypoint detection and meter detection with higher accuracy and less inference time. Specifically, we construct a neighborhood module, which significantly improves the accuracy of keypoint detection, thereby improving the accuracy of meter readings. The prediction generated by NeighborNet are then delivered into the angle reading module to obtain the final meter reading. Experiments validate the effectiveness of our automatic reading method and also prove that our method is robust to different types of meters.

In this paper, we propose a novel detection-free framework for cabinet switch state recognition. Different from prior works which adopt object detection or detection followed by per-switch recognition, our approach processes the image as a whole. Specifically, a semantic segmentation model is used to generate a coarse semantic map as a temporary result, which will be further refined by an object counting module. Moreover, for higher efficiency, we augment the Fully Convolutional Network (FCN) by introducing a hierarchical feature aggregation structure, forming a lightweight yet effective model called HFA-FCN. The experimental results show that the proposed pipeline outperforms those based on object detection, especially in hard cases: images are low-quality, targets are densely distributed or distorted. To the best of our knowledge, we are the first to formulate the task of switch state recognition as a task of semantic segmentation and object counting.

As deep learning continues to mature, researchers have increasingly turned to deep neural network to process object detection in images. While the accuracy of object detection continues to increase, researchers can analyze object features in images better. Image quality assessment is an important research topic in computer vision. When analyzing and processing images, it will lead to more computing overhead if different tasks are performed separately. It can save computing overhead that combining object detection and quality assessment. In order to process quality assessment better while detecting object and to improve the efficiency of image processing, we introduce R-MSSIM, a new system of image quality assessment stronger based on region proposal and structural similarity measure. R-MSSIM can improves the performance of classical algorithms greatly based on object detection. R-MSSIM fits the features of local regions to the final mean opinion score(MOS) while detecting region proposals in an image, which makes the interpretability of image quality assessment stronger. Finally, we prove that our method has a large performance improvement compared with the classical algorithm through a series of experiments, and achieved relatively good performance.

In this paper, morphological operations on regular hexagonal structures are considered, where a regular hexagonal structure is a subset of a hexagonal lattice and consists of the sampled points for the discretization of a regular hexagonal region. First, a sequence of reasonable structure elements (SE) for hexagonal lattices are provided, and the decompositions of the SEs are shown. Second, based on the decompositions of the SEs, some efficient algorithms for morphological operations on the regular hexagonal structures are developed. Finally, the algorithms are tested using a computerized tomography (CT) image, and promising applications of such algorithms on CT image reconstruction and segmentation are pointed out.

In order to improve the accuracy of passenger flow statistics and the adaptability to the environment, the technology of passenger flow statistics based on binocular stereo vision was studied, and a top-mounted binocular passenger flow statistics monitoring terminal was designed and implemented. The research results show that the solution's adaptability to complex environments is much higher than that of monocular passenger flow statistics and other technical means. It can adapt to drastic changes in lighting, complex object interference, and crowded people. The accuracy of passenger flow statistics can be as high as 95% or more; it has realized the height filtering technology, and can customize the height of the people who need to be filtered and counted; using non-contact statistics, the installation of the device is hidden, and it does not affect the feeling of people entering and leaving.

The characterization of a set of target sample can be obtained from a specific sample set. The precision of the model largely depends on the number of the characterization samples and the number of colors. In this paper, a sample selection method combining the third-order root polynomial color correction(RPCC) model is proposed. This method is named Maxmingfc, which is based on the Maxminc method proposed by Cheung et al. And takes GFC evaluation criteria as the criteria for selecting differences between samples, it can use fewer samples to obtain higher characterization model precision. This sample selection method is used to obtain characteristic models under different cameras for different training sample sets and test sample sets, and color difference is used for evaluation. Compared with Hardeberg, Maxminc, Maxmins, Maxsums and Maxsumc methods, the sample selection method of Maxmingfc can obtain the more accurate characterization model, which is superior to other existing sample selection methods.

In view of the technical problems of Beidou / GNSS multi-frequency multi-mode navigation signal compatibility and interoperability and wide area real-time high-precision Lane level positioning, the physical characteristics and difference characteristics of the new generation Beidou navigation satellite signal are analyzed, the deviation identification, modeling expression and correction compensation methods of Beidou/GNSS multi-frequency multi-mode signal are studied, and Beidou / GNSS is established Based on the real-time precise positioning model of wide area, the adaptive fusion and optimization of multi-source and heterogeneous constellation data are studied, and the solution of Beidou / GNSS rapid precise positioning aided by enhanced information of reference station network is proposed. The experimental results show that the scheme can improve the navigation and positioning in the complex urban environment, realize the real-time and fast decimeter level positioning, and the positioning accuracy in the open road is better than 10 cm.

In view of the technical problems existing in the perception and fusion of unstructured data such as audio and video, the method of deep learning is used to realize the pixel level sound source location of video by combining the synchronization of audio and video content in the physical scene and analyzing the sound and image jointly. At the same time, aiming at the problem of low resolution and poor quality of face image in video, a recognition method based on face super-resolution is proposed to realize the identity attribute calculation of the target person.

Face image recognition is an important research topic, which is widely used in many fields. This paper discusses the content of this aspect, briefly analyzes the main technical methods of face image recognition, and expounds and discusses the key points and difficulties in the design, in order to better promote the progress and development of this technology, and make a positive contribution to the related applications in the future.

The low-frequency Synthetic Aperture Radar (SAR) can penetrate foliage and detect the targets concealed under the foliage, while the high-frequency SAR cannot. The low-frequency SAR image and high-frequency SAR image cannot be directly compared to detect the foliage-penetrating (FOPEN) targets due to their distinct statistical properties. This paper presents an FOPEN target detection method based on bi-frequency SAR and conditional generative adversarial networks (CGAN). The high-frequency SAR image is translated into low-frequency one by the CGAN. The direct comparison between the real and generated low-frequency SAR images is used to detect the FOPEN targets. Experiments on P- and Ku-band SAR images show that our method performs better than the double-parameter constant false alarm rate detection (DP-CFAR) using the single P-band SAR image.

Generative adversary networks (GAN) have recently led to highly realistic synthesized image. For the current GAN-synthesized faces detection methods exist false prediction if the real faces with angles or occlusion. This paper proposes a GAN-synthesized faces detection method based on Deep Alignment Network (DAN), which improve prediction accuracy of real faces by makes the locations of facial landmark points more precise. Our method first uses DAN to obtain the locations of facial landmark points of real and synthesized faces; then the landmark points are converted into feature vectors by principal component analysis (PCA); finally, input feature vectors to the constructed Support Vector Machine (SVM)classifier for training. Experimental results show that our method achieves better performance than other method under face with angles or occlusion.

Multimodal medical image fusion is to extract information from different modal images into a single one and obtain the organizational characteristics of source images. Different medical imaging measures such as CT and MRI often bring different visual morphology, however, the salient features of tissues are basically the same from the perspective of human eyes. According to this characteristic, an improved image fusion algorithm based on visual salience detection is proposed in this paper. First, the GBVS algorithm was introduced to calculate visual salience of two registered source images, and then decompose the source images in NSST domain to obtain their low-frequency and high-frequency sub-bands. For the low-frequency sub-bands, local energy and GBVS graph are input into fuzzy logic system to obtain the respective weights for the fused low-frequency sub-band. For the high-frequency sub-bands, the NSML values of each sub-band were calculated and compared to obtain the fused high-frequency sub-band. The final fused image was obtained by using the inverse NSST transformation. Applying this method to multimodal medical image fusion, the visual quality of the image can be enhanced effectively and the salient features of tissues can be preserved well. Experiments on multimodal fusion of different gray-scale medical images show that the proposed method has advantages in retention of image salient features and the overall image contrast, and has better objective index than the comparison models.

In order to improve the accuracy of road surface disease detection, an artificial intelligence-based highway pavement 3D information perception software system scheme is proposed. This article first proposed the hardware composition of the system and the overall system architecture scheme, and introduced in detail the method of implementing high-speed acquisition and storage software system functions. Subsequently, a collaborative control algorithm for coordinating and orderly working the sensors of the three-dimensional road inspection vehicle is proposed, which can avoid conflicts in the work of the sensors. Finally, the specific algorithm flow of the highway pavement 3D information perception software system is proposed. The actual experimental results show that the system can extract the highway pavement 3D information at high speed and accurately.

In order to improve the accuracy of road surface disease detection, the development technology of unmanned aerial vehicle (UAV) in the intelligent transportation field are analyzed. Firstly, the framework of pavement disease recognition and perception system based on UAV is constructed; Secondly, the pavement image data collection experiment is carried out with Yuneec H520 UAV, and the pavement disease image preprocessing technology based on wavelet threshold transform is analyzed; Thirdly, the pavement disease image preprocessing technology based on the DPM is studied, and the recognition method of pavement disease based on VGG-16 neural net-work model is proposed. Theoretical analysis and experimental results show that the method of pavement disease identification based on VGG-16 neural network model has high classification accuracy, and its classification accuracy is better than 90％.

Quality optimization greatly improves coding efficiency by optimizing Quantization Parameter (QP). Recently, preanalysis module to evaluate temporal information propagation inter-frames is used to adaptively adjust QP for each coding unit (CU). However, considering characteristics of Screen Content Coding (SCC), problems exist in module design and QP generation part for screen content. For module design, SCC mainly uses intra-frame compression tools, which haven’t been included in current information propagation model. For QP generation, CUs in fixed-size quantization group (QG) are encoded with same QP, which is imprecise for screen content. This paper proposes intra-frame backwards pre-analysis method and flexible QG size adjustment method for SCC quality optimization. Proposed intra-frame pre-analysis method extends information propagation model by adding intra-frame analysis. Before compression, pre-analysis module finds all intra-frame references and generates importance for each micro block. Proposed flexible QG size is applied during compression and make QG size fit actual CU size. Thus, one QG only contains one CU, so QP of current CU is generated only from the average importance of micro blocks belonging to it, without influenced by micro blocks in other CUs. The experiment shows that proposed algorithm achieves 0.571% BD-rate reduction for the All Intra (AI) configuration with less than 5% encoding time increasing under the SCC common test condition. And the proposed algorithms are conducted in VTM version 5, the standard test environment for Versatile Video Coding (VVC). This work explores the potentiality of quality optimization for screen content and will make contribution to the development of next generation SCC standard.

License plate (LP) location is a key technology in the process of license plate recognition (LPR). How to realize the LP location and character segmentation of complex vehicle images has always been a hot issue in the research of intelligent transportation systems. In this paper, a novel LP location method combining the connected domain (CD) slope detection method and the dynamic template matching method is proposed for the complicated situations that the tilt angle of the LP is too large, the LP is defaced, the LP frame is unclear and the characters are adhesive or defective. Firstly, the method uses the CD slope detection method to find out the equal-slope CD of the LP characters in the optimal segmentation image after pre-processing and determine the tilt angle. Then locate the horizontal region of the LP and realize the tilt correction. After that, the proposed the dynamic template matching method is used to locate the vertical position of the LP and segment the LP characters accurately. Finally, the experiment proves that the proposed algorithm reduces the recognition difficulty of the LP with the above problems, and has the characteristics of fast speed, accurate recognition, good adaptability and strong anti-interference. It also has good versatility and scalability for the newly introduced new energy LP with 8-character.

In order to protect personal privacy, news reports often use the mosaics upon the face of the protagonist in the photo. However, readers will feel uncomfortable and awkward to this kind of photos. In this research, we detect the eye mosaic and try to use eye complementing which is not the same with original picture but matches the nearby texture. It can arouse readers' interest in reading. Traditional in-painting research is not suitable for filling special or large objects, such as eyes or boxes on the ground. They only can fill a small area of missing parts or a single background refer to nearby textures, such as landscape photos. We use WGAN-GP that can refer to nearby textures to generate special objects for in-painting eyes. We also divide the training set into male and female according to gender to avoid eye makeup appearing in all pictures. The experiment result shows our method get higher score in blind testing.

In this paper, we presents a novel image progressive deliver approach based on the continuous relationship, frame prediction and frame buffer between the drawn image sequences, which is a very simple and efficient method, especially when the network is unstable or the network bandwidth is not large. It makes use of the strong continuity between the action relativity in the user's actual operation and the continuous frames in the scene rendering, selects some key frames in a series of rendered high resolution images for progressive network transmission. The client of the network can use the received series of rendered images, then reconstruct the images of other viewpoints according to the images of these key frames, and set the image frame buffer to save the generated or transmitted images, so as to ensure a better user experience when the network is unstable. In the browsing process of 3D scene, it greatly reduces the number of images that need to be drawn by the background server, and also reduces the amount of image data actually transmitted by the network. In this way, under the same network bandwidth, it can smoothly browse the super large 3D scene by multiple people in real time, and improve the user's browsing experience. First of all, our algorithm divides and recognizes the action behavior of the browsing user, transforms the continuous interaction behavior of the browsing user into a series of standard behaviors, which can quantify the interaction behavior of the user, and provides a unified standardized standard for the subsequent prediction, image generation and image caching. Then select the image to be drawn or cached according to the interactive action of the browsing user, and select the image to be sent according to certain rules. Finally, the client program reconstructs the other time according to the received important frame sequence and a series of known constraints, and uses some strategies to cache some frames to prevent network jitter.

Aiming at the hierarchical index method of large image features, the hierarchical index structure of large image is established by using the global position constraint information of large image. This method combines LSH algorithm and KD tree algorithm based on location information. It can control the space complexity of LSH algorithm and the time complexity of KD tree algorithm in high-dimensional feature retrieval. It can speed up the search speed and improve the accuracy of finding the nearest neighbor features, so as to obtain the best comprehensive search performance compared with LSH algorithm and KD tree algorithm. This method mainly includes: extracting thumbnail to build two-layer pyramid image index structure, using LSH algorithm to search rough location for thumbnail, and using KD tree algorithm based on location information to search feature points near neighbor for original image.

Aiming at the hierarchical index method of large image features, the hierarchical index structure of large image is established by using the global position constraint information of large image. This method combines LSH algorithm and KD tree algorithm based on location information. It can control the space complexity of LSH algorithm and the time complexity of KD tree algorithm in high-dimensional feature retrieval. It can speed up the search speed and improve the accuracy of finding the nearest neighbor features, so as to obtain the best comprehensive search performance compared with LSH algorithm and KD tree algorithm. This method mainly includes three steps: extracting thumbnail to build twolayer pyramid image index structure, using LSH algorithm to search rough position for thumbnail, and using KD tree algorithm based on position information to search feature points near neighbor for original image.

The basic KD tree nearest neighbor search has better retrieval performance for low dimension feature space. However, when there is no time requirement to estimate the parameters from the training data, it is difficult to find the nearest neighbor time in the large training set. The number of observations required in the training data set increases exponentially with the increase of dimension. For the 128 dimensional SIFT feature vector nearest neighbor search in this study, simply using KD tree is not the best choice. So we adopt the method of hierarchical index to solve the problem of large image with global plane constraint information. For large original image, we can assist the next index according to the coarse position information returned by LSH index of upper thumbnail. First, for the lower layer, we need to construct an improved KD tree index structure based on location information. The main idea of KD tree index structure based on position information improvement is to establish a plane coordinate system with the upper left corner of the two-dimensional image plane as the coordinate origin, the horizontal direction as the X axis and the vertical direction as the Y axis. In the process of constructing the fourth resolution of the first five nodes of KD tree, X and Y coordinate information are used alternately to split the nodes. In this way, we can roughly locate the feature matching area to the size of the original image, so as to improve the retrieval accuracy and speed at the same time.

The polarization phenomenon of the surface of an object and its changing in different spectra include its surface spatial geometric information and material information. Based on Kirchhoff law and Jones vector , the polarization model of the emission and reflection on the surface of the object is established, and the polarization phenomenon in infrared(IR) and visible light with different materials and incident angles are simulated. The IR and visible binocular polarization imaging system was constructed and the actual polarization data of small unmanned aerial vehicle(SUAV) and buildings were obtained. Two types of characteristic parameters, the degree of polarization and the angle of polarization, were extracted and analyzed, and the results proved that the SUAV and the background of the buildings had obvious differences in IR and visible. This research provides a basis for SUAV target detection and tracking using IR and visible polarization imaging in complex backgrounds.

One common interest in radiography is producing radiographs with as low as possible radiation exposures to patients. In clinical practices, radiation exposure factors are preset for optimal image qualities to avoid underexposures which will lead to repeating examinations hence increasing radiation exposures to patients. Underexposed radiographs mainly suffer from Poisson noises due to inadequate photons reaching the detector. Radiographs are often degraded by scatter radiations and the severity of image quality degradations depends on the amount of scatters reaching the detectors. In this work, a convolutional neural network (CNN) algorithm was used to predict scatters and reduce Poisson noises. Monte Carlo simulation images and an adult abdomen radiograph were used to evaluate this CNN algorithm. The radiograph was underexposed by 60% radiation exposures. The simulation images were produced with one-thousandth of a typical clinical exposure. The results show that Poisson noises are successfully reduced, and image contrast and details are improved. After the underexposed radiograph which is not useful for making a confident diagnosis was processed using the CNN algorithm, the contrast and details in the radiograph were greatly improved and are adequate for making a diagnosis, therefore a 60% radiation dose reduction was achieved. This work shows that radiograph qualities can be improved by reducing scatters and Poisson noises. A potential application of this CNN algorithm is for patient radiation dose reductions by reducing current preset optimal radiation exposures and then using this algorithm to enhance the image contrast and details by reducing both scatters and Poisson noises.

When analysing Internet traffic today it can be found that digital video content prevails. Its domination will continue to grow in the upcoming years and reach 82% of all traffic by 2021. If converted to Internet video minutes per second, this equals about one million video minutes per second. Providing and supporting improved compression capability is therefore expected from video processing devices. This will relieve the pressure on storage systems and communication networks while creating preconditions for further development of video services. Transform and quantization is one of the most compute-intensive parts of modern hybrid video coding systems where coding algorithm itself is commonly standardized. High Efficiency Video Coding (HEVC) is state-of-the-art video coding standard which achieves high compression efficiency at the cost of high computational complexity. In this paper we present highly parallel GPU accelerator for HEVC transform and quantization which targets most common heterogeneous computing CPU+GPU system. The accelerator is implemented using CUDA programming model. All the relevant state-of-the-art techniques related to kernel vectorization, shared memory optimization and overlapping data transfers with computation were investigated, customized and carefully combined to obtain a performance efficient solution across all applicable transform sizes. The proposed solution is compared against reference implementation which uses NVIDIA cuBLAS library to perform the same work. Obtained speedup factors for DCI 4K frame are 2.46 times for largest transform size and 130.17 times for smallest transform size what revealed substantial performance gap of this library when targeting GPU of the Kepler architecture. Achieved processing time of frame transform and quantization are up to 4.82 ms.

The latest video compression standard Versatile Video Coding (VVC/H.266) adopts many advanced encoding tools which achieve larger BD-rate reduction but bring much time consumption. Among all new tools, an inter-prediction mode called InterIMV mode is most time consuming for occupying 12.7% of entire encoding time, with 96.8% of its check redundant in depth level and coding unit (CU) level. Current depth level mode skip methods do early termination based on original top-down structure and skip mode check in lower depth, which brings large quality decrease. For CU level, the common features are not effective to correctly skip InterIMV check. Thus, fast InterIMV algorithm targeted to remove InterIMV mode redundancy is needed. This paper proposes bottom-up InterIMV check method on depth level and temporal rate-distortion relation based check skip method on CU level. Firstly, proposed method executes InterIMV mode in a bottom-up order after all depths are traversed. This method checks InterIMV mode from the lowest depth towards upper depths and finds the suitable depth for current area. Then, InterIMV check in depths above suitable depth is removed without bringing much quality damage, since only the suitable depth is used to encode current area. Secondly, for each CU, the proportion of rate in rate-distortion cost (RD-cost) of temporal best mode is used to judge whether to skip InterIMV check, since this relationship is strongly associated with InterIMV selection. While testing on VVC benchmark VTM5 under RandomAccess configuration, the proposed methods achieve 79.78% speed up in InterIMV process on average. The BD-rate increase of proposed methods is only 0.37% which brings no significant coding efficiency decrease.

In this paper, an infrared image enhancement algorithm based on atmospheric transmission is proposed. First, the algorithm estimates atmospheric light based on a priori principles, that is, the pixel area representing infrared atmospheric light is generally in the upper part of the image and its local contrast is low and its brightness is high. Then, the algorithm estimates the transmittance of the image based on the principle of maximum local contrast of the image, and it uses a guided filtering algorithm to optimize the transmittance. According to the physical model of infrared attenuation in the atmosphere, the infrared image before attenuation is restored. At the same time, this paper uses guided filtering to decompose the image into a basic sub-figure and a detail sub-figure. Among them, the SVD decomposition method is used to improve the brightness of the basic sub-figure. Then, the detail sub-figure and the basic sub-figure are processed by the wavelet transform method, and finally the enhanced infrared image is obtained. Experimental results show that the algorithm in this paper can effectively enhance the infrared image, and the enhancement effect is better than that of ordinary HE, CLAHE, Retinex and other algorithms.

With the rapid development of 3D acquisition technology, 3D point cloud data has become a fast and accurate means to obtain 3D spatial information. However, as a kind of non-Euclidean data, the point cloud's disorder and sparseness make it difficult to train directly using conventional CNN. We propose a 3D point cloud semantic segmentation network based on graph convolution and U-shaped network structure. Among them, the graph convolution module with attention mechanism is used to aggregate the local features of the point cloud data. The U-shaped network with random sampling is used for semantic segmentation to make the network more accurate and more robust. Finally, we evaluated our model on S3DIS indoor data set and got competitive results compared with the benchmark method.

In recent years, there are many applications of object detection in remote sensing field, which demands a great number of labeled data. However, in many cases, data is extremely rare. In this paper, we proposed a few-shot object detector which is designed for detecting novel objects based on only a few examples. Through fully leveraging labeled base classes, our model that is composed of a feature-extractor, a feature attention highlight module as well as a two-stage detection backend can quickly adapt to novel classes. The pre-trained feature extractor whose parameters are shared produces general features. While the feature attention highlight module is designed to be light-weighted and simple in order to fit the few-shot cases. Although it is simple, the information provided by it in a serial way is helpful to make the general features to be specific for few-shot objects. Then the object-specific features are delivered to the two-stage detection backend for the detection results. The experiments demonstrate the effectiveness of the proposed method for few-shot cases.

Aiming at the problems of limited coverage, limited construction and high maintenance cost of the traditional urban waterlogging early warning system, a large-scale waterlogging automatic recognition system based on urban public security monitoring video is designed. By manually organizing the existing urban policing video image data, a set of flood image data set is constructed to train and evaluate a variety of deep convolutional neural networks, and the networks with moderate recognition performance and computational cost are deployed in the actual identification system. The experimental results show that the system can identify the image with 97.72% accuracy, and has a recall rate of 95.16%, and a single image recognition server can support 135 video points, and support elastic scaling, which can meet the needs of large-scale deployment.

This paper presents a local structural feature description of point cloud to efficiently extract local geometric and structure features from LIDAR data for 3-dimensional objective. This approach using hierarchical projection to maps neighbor points with different radial distance to multi-Mercator layers to obtain different distance information of neighbor points to key points. The Mercator projection, a conformal mapping method, the preserves geometric and structure relationship properly. The local features of key points can be obtained by calculating the distribution histogram of each Mercator planes with normalization method. Comparing the proposed approach with other hand-crafted feature extraction methods on Stanford Bologna dataset and 3Dmatch dataset, our methods outperform on descriptiveness, robustness to noise.

The person-related vision tasks face many challenges, such as insufficient or lack of diversity of datasets and learning discriminative identity-sensitive and view-invariant features in the presence of large pose variations. To address this issue, this paper proposes a human posture transfer method based on improving CycleGAN. Transferring a given person’s posture to the target one, while keep the character's identity unchanged, thereby augment the diversity of datasets. The generator of the network contains a sequence of transfer blocks which have similar structures, each transfer block is utilized to transform different part of body as a local transfer. Therefore, avoiding learning the complex structure of the global manifold and address the large spatial misalignment issues induced by transformations of target pose. The discriminator of the network comprised of two convolutional neural networks, to judge the appearance and shape. Quantitative comparisons with state-of-the-art, the proposed method can generate images with the highest score on the metrics, and get performance boost of Re-ID on account of data augmentation.

This article designs smart trash that can be based on machine vision. Its primary function is to complete the automatic identification, classification, and bucketing operations of the typical garbage types in the community. The system uses a Raspberry Pi equipped with a convolutional neural network. It embeds a deep learning model to realize the identification of garbage and online communication to control the mechanical system to complete the class i f icat ion of was te into bucket s . In this paper, the accuracy of the AlexNet model, ZFNet model, and Inception V1 model top-1 are 62.5%, 64%, and 69.8%, respectively. The top-1 accuracy rate of the vgg model has reached 74%, and the accuracy of the model training can be improved to 94% by enhancing the vgg model.

The study explores the approach in business English teaching and learning with the application "51 Learning”, an AI-supported business English training platform to cultivate the ability of active knowledge construction, critical thinking and analysis, creative problem solving and business practice and operation, thus to achieve the purpose of deep learning. AI is the technical means, blended learning is the teaching means and method, and deep learning is the goal. The study shows it is of great benefit in improving practical ability and enhancing theoretical concept in business English; it broadens the channels of business English learning, improves their interest in business English because of its autonomous learning atmosphere, and enriches the evaluation methods and records the whole process of learning. Therefore, the next step of the study is to conduct an empirical research to verify the effectiveness of deep learning in the context of AI-based blended teaching.

To solve the problem of poor sample diversity of infrared ship image data, a method of infrared ship image data expansion is implemented based on CycleGAN. The method using the ideas of circulating, iterative approximation, there can be no noise of ship target infrared simulation image dataset maps to conform to the actual infrared detection scenario of ship target image dataset, according to the demand of the subsequent target detection identification, on the premise of the ship target itself form unchanged, injected with appropriate clutter interference, so as to realize the effective expansion of the image data, the method fully considers the target and background infrared characteristics, and are not influenced by whether the scale of the target alignment issues, which can effectively increase the image sample of diversification, for subsequent ship target detection identification algorithm, provides rich data support. The comparison experiment of image structure similarity and object detection accuracy verifies the effectiveness of the algorithm.

Computed tomography (CT) images-based early disease screening is of high significance for the detection of the occurrence of cancer. It has been demonstrated that lesion detection using deep learning in CT images are significantly effective for the early stage of cancer. In this study, an improved Cascade R-CNN is proposed. In the proposed network, the Feature Pyramid Network (FPN) is introduced to complete automatic computer-aided lesion detection. Based on this trick, numerous tiny lesions in CT images can be well detected. Experimental results on the DeepLesion show that the proposed method can achieve the mAP of 0.598 with a threshold of 0.5 and 85.2% sensitivity with 4 false positives per image.

Human-object interaction (HOI) detection task is defined as inferring all the < human, verb, object > triplets in the image, which helps computers to obtain a more comprehensive understanding of the visual scene. Most existing HOI detection methods focus on instance local features, and rarely consider the information from backgrounds. Our core idea is that the relationship between human, object and other backgrounds contains important cues to facilitate HOI detection. According to the short-term memory selection (STMS) mechanism, we regard the interaction relationship as the result of human and object stimulating the union area, and simulate the stimulation process by the recurrent neural network. The features in the union area of human and object are taken as the input of RNN, human and object are the two inputs of RNN, and the output is the representation of the interaction relationship. Combined with the visual features and spatial features of instances, a multi-stream network is utilized to detect HOIs in the image. Experiments on V-COCO and HICO-DET show that the proposed model achieves better performance, verifying the effectiveness of our method.

Clothing detection and landmark detection are important techniques in fashion image analysis. The availability of large annotated fashion datasets has made fashion image analysis a hot research topic. This paper proposes a single-stage detector that performs bounding box detection, fashion landmark detection, and can also predict end-to-end clothing category classification. This parallel processing provides improved time efficiency than the later technique that performs regional proposals first and then prediction module. The proposed network is designed with the revision of the EfficientDet model announced by Google Brain. The proposed approach can also be used within a real application because it can operate efficiently and quickly from the inference latency perspective.

In recent years, due to the rapid development of deep convolutional neural networks, deep learning model inference needs to consume a lot of computing resources. Most current edge devices cannot support deep learning applications with low latency, low power consumption, and high accuracy due to limited resources. Deep learning applications. Therefore, model compression and acceleration of deep networks are an effective solution, and network pruning that simplifies the model by removing redundant parameters in the inference stage is a hot research in this field in recent years. This paper divides the work into six aspects for a detailed analysis, combs the latest progress of deep neural network pruning technology from the perspective of different granular pruning and weight measurement standards, and finally points out the problems in the current research and analyzes Future research directions in the field of pruning.

In low textured environments, visual Simultaneous localization and mapping (SLAM) only using point features is challenging in tracking sufficient point features, and this leads to poor accuracy and robustness. For this problem, this paper proposed an improved visual-inertial odometry (VIO), which can make better use of the advantage of point and line segment features to achieve fast and robust results. Visual features including points and lines are extracted and tracked using optical flow method. For line segment features, Plücker coordinates are employed for 3D spatial line and optimized using orthonormal representation. Furthermore, point features and line endpoints are expressed using inverse depth. To fuse the data from visual sensors and inertial measurement units (IMU), the states including IMU states and 3D landmarks are optimized by minimizing the measurement residuals including visual re-projection error and pre-integrated IMU error in a sliding window. Our proposal is evaluated on EuRoc datasets and compared with state of the art methods. Our method can achieve more robust performance in most of the experiments while maintaining its speed and accuracy.

6D object pose estimation is a fundamental problem for many computer vision and robotics applications. Recent work has shown that data-driven approaches could enable accurate 6D pose estimation for objects with sufficient texture on the surface. However, few works have focused on estimation 6D pose for texture-less objects. In this paper, we present a network that estimating 6D pose for texture-less objects by using the multi-scale relational features. The proposed network, which leverages both the appearance and geometry features from multi-scale point groups, is able to extract distinctive features for texture-less region. In particular, the multi-scale features encode relational information of the point groups, are more informative compared to the feature comes from vanilla convolutional neural networks and PointNet. The proposed network is end-to-end trainable. Experiments on T-LESS dataset demonstrate our method achieves competitive results on 6D pose estimation task of texture-less objects.

Orientation selectivity of neurons in mammalian visual cortex is a unique physiological characteristic. In the primary visual cortex, neurons with the same orientation cluster together to form the orientation column of the visual cortex. The mechanism of orientation selectivity has been debated for a long time, and the discussion of the structure and function of orientation columns has also been a hot issue. Here we present two basic properties of the visual cortex response, orientation selectivity and surround suppression. Based on these two different characteristics of primary visual cortical neurons, an improved EN model is proposed. The model can well simulate the physiological structure of simple cells and take into account the suppression and facilitation of non-classical receptive fields.

Aiming at the problems of large dimension and low recognition efficiency of human action sequence recognition, a recognition algorithm based on correlation feature selection and multi-layer perceptron is proposed. Use Kinect to obtain continuous images of human movements. By extracting key frames of human movements and calculating human joint angles, key features of human movements are extracted. Through the correlation feature selection algorithm, the most effective feature subset is selected from a group of features to reduce the dimension of feature space and reduce the calculation amount of classification recognition. Finally, the multi-layer perceptron is used as the classifier, and the BP algorithm is used to adjust the network weights to realize the classification and recognition of human action sequences. Through the classification and recognition experiments of human action sequences, it is proved that the recognition algorithm based on correlation feature selection and multi-layer perceptron has a high recognition rate. The recognition rate of the algorithm proposed in this paper is as high as 96.667%.

Synthetic aperture radar (SAR) is an active microwave imaging sensor, which can provide images in all-weather and all-day conditions. The various scales and irregular distribution of different ships in SAR images, is a heated and challenging problem. As a basic component in the object detection frameworks, Feature Pyramid Networks (FPNs) improve feature representations for detecting objects at different scales. However, FPN adopts the same convolution operation at different layers, which does not consider the differences between different levels. In this paper, we present Dense Feature Pyramid Network (DenseFPN). Based on the hierarchy of backbone network, the cross-scale connections and lateral connections, the shallow features and deep features are processed differently in DenseFPN. Compared with conventional FPN, we integrate DenseFPN into Faster R-CNN framework and thus form a novel detector. Experiments on high-resolution SAR images dataset (HRSID) have verified the effectiveness of the enhanced hierarchical feature in the proposed method compared with other typical CNN based methods.

Federated learning is an emerging machine learning setting, which can train a shared model on large amounts of decentralized data while protecting data privacy. However, the communication cost of federated learning is heavy, especially for mobile devices with higher latency and lower throughput. Although several algorithms have been proposed to reduce the communication cost, they are extremely sensitive to data distribution, even inapplicable to the real client Non-IID data. In this paper, we propose an effective communication strategy for federated learning called FedSAA, which increases the testing performance on Non-IID data by introducing self- attention mechanism. Two major innovations of our paper are presented here. Firstly, we utilize self-attention mechanism to optimize both the server-to-client and the client-to-client parameter divergence during the model aggregation process so as to improve the model robustness for Non-IID data. Secondly, we adopt the sign compression operator to help data transmission between nodes. The experimental results demonstrate that the model accuracy of our communication-efficient strategy for federated learning with Non-IID data is superior to other communication-efficient algorithms.

Event extraction is a challenging task in text mining. Current research on event extraction is often based on the sentence level, while the research based on the document level is less investigated. In this paper, we propose a deep learning model, termed BLMA, which combines Bi-LSTM and Multi-head weighted Attention to detect multiple event types. Unlike traditional methods, BLMA can directly detect events in texts without trigger words. We regard the event detection task of the chapter as a multi-label classification problem and apply the BLMA onto the application of detecting events in movie scripts. Experiment results show that BLMA performs well on an automatically tagged screenplay dataset containing 11 event types. F1 score of event detection using BLMA reaches 91.6% for a chapter with a single event, and 93.9% with multiple events, which outperforms other models. Therefore, our deep learning approach can effectively extract events from a chapter, which can be used for event detection on movie scripts. In addition, we explore the contribution of each component of the BLMA to the results, and find that multi-head attention has a significant impact on the performance. Keywords: Event Detection, Bi-LSTM, Attention.

This paper proposes a learning model of basketball players structure based on the Bayesian network. By analyzing the data of NBA 's players, we complete the structural learning of basketball players network based on the +/- values of 5 resident players in the Portland Trail Blazers team. We finally obtain a winning and losing models for the five resident players of the Portland Trail Blazers team, and we make suggestions for coaches about player rotation based on the analysis of the models.

The computer network has been widely used in various industries of society, and network security has received unprecedented attention. Network intrusion detection technology is the critical technologies, which can maintain network security. However, the traditional rule-based intrusion detection method has some shortcomings, such as relying on manual intervention, and it is difficult to update the rule database in real-time. Therefore, in this paper, we propose a novel network intrusion detection model based on deep attention neural network. In particular, we combine the LSTM, multi-layer perception and the attention mechanism in an end-to-end model in order to extract features automatically by deep learning technologies. Finally, we conduct extensive experiments on the KDD99 and NSL-KDD dataset, and the results demonstrate the effectiveness of our proposed approach.

The automatic identification of overpass structures is of great significance for multi-scale modeling, spatial analysis, and vehicle navigation of road networks. The traditional method of overpass recognition based on vector data relies too heavily on the characteristics of manual design and has poor adaptability to complex scenes. In this paper, a method for overpass identification based on the target detection model Faster R-CNN (Regions with Convolutional Neural Network) is proposed. This method uses a Convolutional Neural Network to learn the deep structural characteristics of data samples, and then automatically identifies and finds accurate positioning of the overpasses. The experimental results show that this method is able to identify overpasses and can accurately determine their positions in a complex road network, avoiding the influence of human intervention on the uncertainty of results. This method also has strong anti-interference abilities

Loan default prediction has been playing a key role in credit risk management throughout the years. Existing solutions usually involve classical machine learning classifiers, e.g. logistics and SVM, but most of them need extensive feature engineering such as feature cross which requires plenty of hand-crafted feature design. In this paper, we propose a novel method to implement feature cross based on the convolutional neural network. This method is designed to extract automatically important cross features and generate cross-feature embedding from structured data which reduces the need to generate hand-crafted cross features. The experimental results show that our method can improve the performance of predicting loan default probability compared with the methods based only on classical machine learning algorithms that are widely used in loan default prediction.

Recent studies demonstrated the clinical utility of seismocardiography (hereinafter SCG) signals for the detection and monitoring of cardiovascular conditions. Renewed interest in investigating the utility of SCG has been accelerated recently and benefited from new advances in low-cost lightweight sensors and machine learning methods. This article compares various machine learning algorithms (the method of nearest neighbors, the method of support vectors, decision trees, the ensemble of models) and neural networks: based on the architecture of long short-term memory and convolutional ones. An original numerical experiment was carried out using the developed mathematical software, where all of the mentioned methods and algorithms were implemented. During this study, much attention was paid to the preparation and preliminary processing of data. In particular, signal filtering is carried out using the Butterworth filter, and the issues of extracting features from the signal, which will become an input vector for machine learning algorithms, are also discussed. To compare the effectiveness of the considered models for solving the problem of diagnosing diseases, Accuracy, Recall, Sensitivity, Specificity, Precision, F1-measure, etc. are given. For each algorithm and data set, confusion matrices and ROC curves were constructed. Results of this research show that convolutional neural networks are very effective at diagnosing the states of the human cardiovascular system and supporting decision-making in cardiology and cardiac surgery.

Heart rate variability (HRV) refers to the variation of the heart rate cycles, which contains information of how the autonomic nerves system regulates the cardiovascular system. HRV is a valuable indicator to diagnose various cardiovascular diseases and predict arrhythmia events. This study is based on the standardized five-minute and ten-minute RR interval series from the open source Electrocardiogram (ECG) database website PhysioNet. Artificial Neural Networks (ANN) are used to distinguish patients with congestive heart failure or atrial fibrillation from normal sinus rhythm utilizing features calculated by time and frequency domains as well as nonlinear analysis. To eliminate redundancy and avoid overfitting, Principal Component Analysis (PCA) is performed to screen for the most efficient features. PCA not only improves the accuracy but also greatly reduces the number of nodes in the ANN model, thus, improves the efficiency. Overall, ANN classifiers achieved an accuracy of 79% for five-minute RR interval series and 84% for that of ten-minute series. The performance of Random Forest (RF) classifier is not as satisfactory. However, its list of most important features indicates nonlinear dynamics may play an important role and provide useful insights to the classification problem.

In order to solve the problem of MIMO radar underdetermined blind separation under the condition of overlapping in time domain, frequency domain and time-frequency domain, a blind separation algorithm based on sparse smooth reconstruction and ridge estimation is proposed. Firstly, underdetermined blind signal separation is simplified to signal reconstruction with known mixed matrix. Secondly, using the time-frequency sparsity of MIMO radar signal, a time frequency smooth sparse reconstruction algorithm is proposed, and the time-frequency information is modified by the ridge estimation method. Finally, the time-frequency source signal is recovered by inverse transformation. The algorithm can reconstruct the coding information of source signal and radar signal at the same time, which provides a new way to solve the problem of radar signal separation in complex electromagnetic environment.

Prediction Intervals (PIs) provide a method to quantify the uncertainty of deep neural networks' point forecasts. High-quality PIs should be as narrow as possible while covering a designated percentage of data points, which is an axiomatic theory. Lower Upper Bound Estimation (LUBE) method is the first to incorporate this axiom into the neural networks loss function, however it pays much attention to the interval width, and thus the coverage probability does not achieve the desired result. Consequently, the PIs are unreliable and their practical application risks are elevated. In this paper, prioritizing coverage probability, we propose a coverage-driven approach that is generalized to any neural networks model, combining bootstrap method with improved LUBE method. We show that PIs constructed by our method are more reliable, and model uncertainty is quantified using bootstrap. Moreover, compared with two novel PI methods, benchmark experiments show our method is able to reduce the mean PI width by more than 7.5% while obtaining the better results in coverage probability.

Federated Learning (FL) is a client-based distributed machine learning framework, which aim to train a central- ized model with protecting data privacy. However, the decentralized datasets pose a challenge on the traditional FL, as they are non-independent and identical distributed (non-IID). Non-IID settings can result in client result gradient biases, which may decrease the accuracy of the model. To address this issue, we propose Federated Learning with Auxiliary Generator(FedGen), which keeps the consistent of data distribution between clients leveraging the auxiliary generator, and the gradient become more accurate. To demonstrate the effectiveness of proposed method, extensive experiments are conducted on the benchmark datasets, including the MNIST and LEAF dataset. The experimental results shows that FedGen converges 1.2 times faster than FedAvg, while the accuracy can be increased.

The heterogeneity of geospatial information will persist for a long time. As the key to overcome the semantic heterogeneity, categories mapping has gained considerable attention. In previous studies, the existing geographic ontologies cannot support enough multi-semantic extension to conquer semantic heterogeneity effectively. When introduced to explore categories mapping, many semantic similarity measures encountered the problem of subjective weight setting and failed to make full use of the organizational structure information of categories. The rapid development of Artificial Intelligence (AI) and Natural Language Processing (NLP) bring new enlightenment to the semantic analysis and understanding in the geographic information field. Therefore, this paper proposes a new geographic categories mapping method based on ontology attribute characteristics learning, which utilizes ontology attributes and the classification hierarchy of geographic categories. Firstly, a basic semantic framework based on ontology attributes is defined to realize the semantic vectorization descriptions of geographic categories, by extracting semantic knowledge from definitions. Then, a new hierarchical coding method is proposed to describe the classification hierarchy of categories and identify the classification status of each category. After that, a self-learning mapping mechanism based on BP neural network is used to establish the non-linear relationship between ontology attribute eigenvectors and classification states, which can support categories mapping. Finally, some categories mappings are formed by this method to evaluate transition effects, and introduces the category differentiation degree to analyze the influence of classification structure on prediction accuracy. The preliminary results show the feasibility and reliability of the proposed model for automatic semantic mapping.

Deep reinforcement learning has greatly simplified visual navigation by utilizing the end-to-end network training strategy. Unlike previous navigation methods which build upon high-precision maps, deep reinforcement learning-based method enables real-time navigation by only taking one image as input at a time. As such, deep reinforcement learning based navigation methods are applicable to a variety number of applications in robotics/vision communities, thanks to its light-weight computational cost. Despite the advantages, however, these methods still suffer from inefficient data exploration and poor convergence on network training. In this paper, we propose to use inverse reinforcement learning to solve the problem，which can provide more accurate and efficient guidance for decision-making. The proposed method is able to learn a more effective reward function from less training data. Experiments demonstrated that the proposed method achieves a higher success rate of navigation and produces paths that are more similar to the optimal ones compared to the reinforcement learning baselines.

As a major health problem, traumatic brain injury has received increasing attention in recent years. Mild traumatic brain injury (mTBI) is the representative of the vast majority of traumatic brain injury. However, most studies have focused on moderate to severe traumatic brain injury. According to reports, mTBI patients during the acute period have the most prominent cognitive disorders. Hence, we collected the functional magnetic resonance imaging data in resting-state (rsfMRI) of the acute-mTBI patients and the healthy controls to explore the differences of brain functional connectivity between them. We chose dynamic functional connectivity in resting-state as features and performed feature selection/extraction. Finally, the classifier based on machine learning methods achieved a good classification accuracy of 85.48%. Most of the extracted brain regions of interest (ROIs) with high identification power belong to the sensorimotor and functional connectivities extracted within the sensorimotor has the highest proportion which suggested the sensorimotor may be the most severely damaged part of mTBI patients in the acute stage.

Entity extraction is an attractive but challenging research filed. In recent years, the development of deep learning has promoted entity extraction techniques. However, owing to underutilizing multi-grained context, existing entity extraction methods obtain limited performance. Actually, cross-sentence information is capable of relieving the problem of ambiguity in entity extraction, especially for Chinese language. Instead of exploiting single intra-sentence context, we propose a cross-grained context guided Chinese entity extraction model by mining and integrating inter-sentence, intra-sentence and document-level information. Specifically, a document graph is constructed to simultaneously represent cross-grained information for a Chinese article using dependency syntax. Certain rules are designed to link syntactic trees of sentences to a weighted graph. Then an entity extraction model is established using GCN network to capture deep relations between words on the graph. Finally, cross-grained information embedded by GCN is used to predict entity mentions via multiclassification. Experiment results show that our method has achieved comparable performance to two baselines on Chinese discourse-level literature datasets, with average accuracy, recall, and F1 scores of 79.02%, 67.41%, and 78.08%, respectively.

Deep learning based algorithms have made huge progress in the field of image classification and speech recognition. There is an increasing number of researchers beginning to use deep learning to process electroencephalographic(EEG) brain signals. However, at the same time, due to the complexity of the experimental device and the expensive collection cost, we cannot train a powerful deep learning model without enough satisfactory EEG data. Data augmentation has been considered as an effective method to eliminate this issue. We propose the Conditional Wasserstein Generative Adversarial Network with gradient penalty (CWGAN-GP) to synthesize EEG data for data augmentation. We use two public neural networks for a motor imagery task and combine the synthesized data with real EEG data to test the generated samples’ data enhancement effect. The results indicate that our model can generate high-quality artificial EEG data, which can effectively learn the features from the original EEG data. Both neural networks have gained improved classification performance, and the more complex one has obtained more significant performance improvement. The experiment provides us with new ideas for improving the performance of EEG signal processing.

With the development and maturity of object detection techniques, more and more researchers begin to utilize deep learning methods for object detection and classification of the image data obtained by dynamic vision sensors(DVS). Considering that the event stream data obtained by DVS does not have grayscale feature information, we want to convert it into a frame image and utilize the YOLOv3 neural network model for learning to achieve object detection. Since the IoU loss function in the original YOLOv3 network cannot represent the distance between the output predicted box and the grounding truth, this paper improves YOLOv3 by employing the GIoU loss function to achieve more accurate detection. By conducting experiments on a self-collected dataset, we find that the GIoU-improved YOLOv3 network has good performance and can accurately achieve human action recognition and classification.

Intelligence of weapon system, precision of striking weapon and diversification of combat formulation are important developing trends in future combat operations. A terminal guidance law based on reinforcement learning (RL) and deep deterministic Policy Gradient (DDPG) is proposed to solve the problems in missile guidance system, such as heavy dependence on simulated environment training, poor interception effected by condition constraints, and insufficient guidance accuracy caused by the difference between simulated environment and the real environment. Taking advantage of the maneuverability of the small anti-ship missile relative to the ship, according to the game theory in reinforcement learning combined with the deep neural network, the optimizing trajectory is updated by analyzing the projectile motion and line of sight angle. Reward was set reasonably in line with the spatial position and the strategy gradient was optimized by using the deep neural network. Simulating experiments show that after iterative training, the DDPG guidance model for small anti-ship missiles can optimize the ballistic curve, and the miss distance can meet the requirements. Compared with traditional guidance laws, this model has better autonomous decision-making ability and strike capability.

Identifying the financial fraud behavior of listed companies in a timely manner and helping investors avoid risks is a key measure to promote the healthy development of the capital market. Aiming at the lag of traditional fraud identification method, a text classification model suitable for Chinese capital market is constructed based on user generated content (UGC) on social media platform. To detect fraudulent companies,TF-IDF features, topic features and explosive news quantity features are extracted guided by systemic functional linguistics (SFL) theory. Empirical analysis is conducted by crawling the news and comments of 124 companies from Eastmoney and JRJ.com.The empirical analysis results show that the model constructed by the article can effectively extract the implicit information in unstructured data and improve the timeliness of screening fraudulent behaviors of listed companies. The topic and keyword features in text comments play an important role in distinguishing the fraud of listed companies. On the one hand, it helps individual and institutional investors avoid investment traps, on the other hand, it helps regulators detect companies with fraudulent potential in time to prevent market risks.

Incremental capacity (IC) analysis is an effective and widely used approach to evaluate the lithium-ion battery remaining capacity. A limitation for IC based capacity estimation is that it requires a consistent constant charging current for each test and demands a large amount of data to be stored, which hinder its use in practical conditions. To address the issues, this paper proposes an online capacity estimation method for battery module with incremental capacity curves processed by tracking differentiator. Specifically, the IC features, such as peak voltage, peak amplitude, are analyzed for the battery module empirically and theoretically. The analytic linear relationship between the IC features and capacity are formulated. As a basis, the calculation of dQ/dV is converted to the derivative of terminal voltage, which is obtained by a designed tracking differentiator. For the iteratively derived IC curve, a realtime peak detection algorithm is proposed to detect the second peak voltage without the requirement to store any historical data. The proposed method is not sensitive to noise and applicable to common charging conditions. Experimental results with cycle data for a commercial 206Ah battery module reveals the superiority of the proposed method.

With the popularity of the Internet and the rapid development of information technology, it has become critical to accurately predict the security situation of the network environment and timely keep watch on those potentially dangerous attacks according to the security situation prediction. Therefore, in this paper, we propose a novel network security situation prediction model based on temporal deep learning. We combine the attention mechanism with recurrent networks to learn the historical time series network data's hidden features. Then a predictive layer is applied to analyze the hidden features and predict the network security situation. Our experiment results show that our proposed model is significantly better than ARIMA, DNN, and other comparative models, demonstrating the effectiveness of our proposed model in network security situation prediction.

To address the time-consuming feature extraction and model training in the process of EEG emotion recognition, this paper proposes a method to rapidly train deep learning models for EEG emotion recognition with high accuracy and excellent performance. The DEAP EEG data set is used to quickly train and fit the deep learning model, so as to establish a new pre-trained model for EEG emotion recognition. In addition, it was found that the best training effect was achieved using a sample with a ratio of 25%, and the other test data could quickly fine-tune the original model. The experimental results proved the effectiveness of the method, and the accuracy of the pre-trained model could reach the highest 93.72% in the Valence emotion dimension.

The carbon emissions in the production process consist of direct carbon emissions and indirect carbon emissions. The carbon emissions involved in the CNC centerless grinding process are composed of energy, resources, and waste. In order to illustrate the impact of the carbon emissions during the CNC grinding process. In this paper, through the combination of theory and experiment, we analyze and test the power consumption, grinding fluid consumption and lubricating fluid consumption of CNC centerless grinders respectively, so as to establish an ERWC model that reflects the carbon emissions during CNC grinding.

Scenario definition for Automated Driving Function Validation is the state-of-the-art research area for autonomous driving, unification of the scenario description methods and language is the crucial step for the scenario definition. Based on the simulation scenario database established by China Automotive Technology and Research Center (CATARC) using natural language processing methods, to connect the human readable scenario definition with programmable scenario coding, provide the bridge between scenario definition with simulation testing, for better unify between scenario definition and virtual scenario building. With the participation of many enterprises, the corresponding test of automatic driving has become an access condition for the implementation of the corresponding technologies of automatic driving. Enterprises and institutions all over the world are trying to establish a unified scenario database to achieve the wide sharing and unification.

Trajectory distribution of vessels shares a large difference between coastal waters and exposed waters, thus a segment algorithm is needed when mining the waterway patterns to analysis the traffic flow. Considering traditional spatiotemporal trajectory cluster algorithm cannot meet the requirements on big data processing speed, this paper proposes a wavelet clustering-based vessel channel pattern extraction algorithm. Firstly, a marine navigation feature space is built by means of dividing the target area into grids, which is transformed into a transform domain using twodimensional discrete wavelet transform algorithm in the next step. Then we select an appropriate threshold to find the connected region and sub-regions with different density distributions are obtained. Drawing on the idea of image processing, this algorithm takes the channel boundary and distribution matrix as the overall channel characteristics. In the end, real historical AIS data is used to verify the method. The results show that such can reduce the processing complexity of spatiotemporal trajectory data, and greatly shorten the response time to find vessel channel patterns.

Drug crimes seriously endanger human health and social peace. Strengthening the correlation analysis of the features of drug-related cases and fully mining up the correlation rules from the features of the cases have become the key links for effectively preventing and responding to drug crimes. After performing performance analysis on the two association analysis algorithms, in view of the lack of existing association analysis research and association rule mining methods in drug-related cases, this paper proposes a method for analyzing association rules of drug-related cases, Through the FP-growth algorithm and K-supported expectation method to calculate the weighted frequent items of drug-related personnel, and use the Apriori algorithm to obtain the feature association rules. The results show that the proposed method makes a good trade-off between the amount of data and the minimum support (called minsup), and can effectively mine association rules from the features of drug-related cases. It is of great significance to prevent and control the development of drug-related crimes in time and improve the ability of public security departments to deal with drug-related cases.

Ant colony optimization (ACO) algorithm is an effective method for solving combinatorial optimization problems. However, the ant colony algorithm has the tendency to fall into the local optimal solution, especially for the complex traveling salesman problems (TSP). Aiming at the disadvantages of ant colony algorithm, an ant colony algorithm based on multiple state transition operators (STOACO) is proposed in this paper. First, we select suitable cities from the unvisited cities as elements of the dynamic search table and the dynamic search table is established to enable the ants to find a good solution at the initial stage, improving the convergence of the algorithm. Then, in order to avoid the algorithm falling into the local optimum, three specific state transition operators are introduced to further strengthen global search ability. Finally, STOACO is compared with other ant colony algorithms, and the simulation results show that STOACO has better performance on the TSP.

In some large-scale industrial production processes, when a fault occurs in a unit, it will spread through the connectivity between the units, which can affect the entire factory and cause product quality deterioration or even more serious problems. It is important to diagnose and isolate the root cause. Granger causality analysis is widely used which provides an effective way to localize root cause of faults. However, the conventional Granger causality is not suitable for nonlinear and high-order signals. This paper proposes an effective model-free, copula-based Granger causality method for root cause diagnosis of plant-wide oscillation which can effectively reveal the nonlinear and high-order causality. Granger causality is transformed into log likelihood ratio of conditional distribution and conditional copula is used to derive an effective estimation. The numerical simulation case prove the validity of Granger causality analysis based on copula function. At the same time, for the root cause diagnosis of the actual plant-wide oscillation, this method also successfully detected the correct root cause.

In the era of big data, electronic health records contain a lot of information that can be used, diagnosed, predicted and medicated, but also contain a lot of private information. Therefore, it has received attention from all sides. This paper takes the privacy protection of electronic health records as the main research object, and tries to find solutions to the privacy problems existing in the process of electronic health records management in China, hoping that it can provide some valuable references for the smooth development of electronic health records management in China. The main research content of this paper can be divided into three parts. Firstly, the definition of the basic concept of electronic health records and the clarification of its connotation and composition. Second, the privacy risk analysis of electronic health records, mainly from the personnel level, technical level as well as the legal and management level to analyze the current state of privacy protection of electronic health records and its shortcomings in privacy protection in China. Third, the privacy protection of electronic health records, this part mainly proposes recommendations for improvement based on the problems existing in the privacy protection of electronic health records.