This PDF file contains the front matter associated with SPIE Proceedings Volume 13156, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

With the improvement of computer computing power, the object detection algorithms based on deep neural network has ushered in vigorous development, and has been widely used in industry, agriculture, medicine, military and other fields. One-stage object detection algorithms shows the superiority in real-time detection compared to other object detection algorithms such as two-stage object detectors or ViT-based detectors. At the same time, more and more anchor-free detectors show the advanced nature of anchor-free algorithms compared to anchor-based detectors. In this paper, we review the one-stage anchor-free real-time object detection algorithms in recent years, and analyze the application scenarios and optimization strategies of future object detection algorithms. Firstly, the principle and advantages of anchor-free object detection algorithms and one-stage object detection algorithm are introduced. Secondly, the network structure and innovation of anchor-free object detection algorithms in recent years are summarized. Finally, the possible development direction and trend of one-stage anchor-free real-time object detection algorithms in the future are proposed.

The high-temperature plume of rocket engines has become the target of detection and tracking because of its significant infrared characteristics. The engine type spectral parameters have a decisive influence on the infrared characteristics of the engine tail plume. This paper introduces polynomial chaos and non-intrusive methods to analyze the influence of the uncertainty of the engine type spectral parameters on the uncertainty of the infrared characteristics of the plume through a surrogate model . The results show that the spectral radiance value at the 2.7 μm band is larger, the uncertainty band is wider, and the uncertainty reaches about 25%; the spectral radiance value at the 4.3 μm band is smaller, the uncertainty band is also narrow, and the uncertainty is not high, which degree of certainty is about 10%. However, a peak and a wider uncertainty band appeared in the 4.50-4.75 μm band, and the uncertainty reached 30%.

In response to the imperative requirement of the Background Oriented Schlieren (BOS) imaging system to produce exceptionally realistic target schlieren images, this paper proposes a method for generating realistic background schlieren images based on physical optical tracing. This method, grounded in real-world scene images, couples atmospheric turbulence field information with target information to establish a physics-based background schlieren imaging system ray tracing full-link simulation model. It simulates the propagation of light rays in complex refractive index fields of three-dimensional scenes and calculates the intersection points of light rays with objects based on physical principles. By using diffraction imaging principles, it accomplishes forward mapping of background images, thereby generating background schlieren images with higher realism and credibility. Experimental results demonstrate that this method offers high simulation flexibility and generates background schlieren images that are realistic and reliable, effectively showcasing schlieren characteristics. This method provides essential support for the pre-design and evaluation of background schlieren imaging system systems.

Infrared and visible images have their own characteristics and different imaging principles, the targets be detected by any kind of sensor have their own advantages and disadvantages. In this paper, the characteristics of infrared and visible light images in dim and small target detection are analyzed, and pseudo-color infrared and visible light image fusion detection methods of dim and small target based on visual attention is studied, and its fusion detection effect is analyzed. The research shows that the fusion algorithm in this paper is more effective for targets with complex background and low contrast, and this method only makes pseudo-color mapping on a local area of the target, which is not easy to cause visual fatigue.

Detection, tracking, characterization and identification of space targets orbiting the Earth is one of the main tasks of the space surveillance system. Spatial target image feature extraction can find out some descriptions that reflect the essence of the target from the original image. The moment feature describes the statistical characteristics of grayscale by the moments of each order of image distribution, which can effectively reflect the shape information of the target. A Hu extension moment is proposed, and a feature extraction method for spatial target images is proposed based on this. It solves the problem of information loss when the confidence variance difference is large and the eigenvalue is negative. A linear combinatorial feature calculation method is proposed for spatial target classification, and a linear classifiable quantitative discrimination method is given. The image needs to be preprocessed. Then, the seven eigenvalues based on the extension moment are calculated, and the similarity analysis of the Hu extension moment is carried out. At last, the similarity of the sample target region matching is analyzed, and the target similarity can reach 90%. For image sets with multiple types of targets, according to the numerical distribution calculates the linear combination feature, linear classifiable quantitative discriminant results can be calculated. 48 images of four satellites in the database were selected as test data. Its recognition accuracy rate reaches 95.8%. Calculating the recognition result of the same platform for multi-sample joint recognition, and the probability of recognition attribution after multi-sample union is 100%.

When using Geiger mode Avalanche Photo Diode (Gm-APD) array lidar for long-distance imaging, the few echo photons make it challenging to get the target position. To solve these problems, this paper proposes a spatial correlation extraction algorithm combined with morphological filtering (SCMF), which uses the spatial correlation of the target to superposition the weight of the pixel histogram, increasing the number of statistical frames, improving the signal-to-noise (SNR) of pixel statistical data and accurately extracting the distance value of the target pixel. Spatial correlation also improves the real-time imaging of the system. According to the time-domain spatial dispersion characteristics of residual noise pixels of small intensity threshold, a local spatial distance correlation logic method is proposed, which only preserves the pixel groups with similar spatial distances and removes the stray background noise pixels. Because the number of pixels in the target pixel group is more than the noise group, a spatial filter module is constructed using morphological filtering to remove the remaining blocky noise group and preserve the target pixel group. The proposed method can achieve long-distance imaging in 0.02s acquisition time through outdoor real imaging experiments. Under the echo condition of 0.0152 Signal to Background ratio (SBR), the SCMF method has 76% target restoration, and the reconstructed image SNR can improve 23 times compared with the peak-picking method, a great improvement has been made in the reconstruction of image denoising.

In order to reduce the impact of cloud clutter on target detection in infrared polarization images under sky background, a target detection method based on maximally stable external regions (MSER) is proposed. Firstly, subtract each pixel of the infrared polarization image from the mean of the vertical and horizontal directions of the image. The non positive pixel is the cloud background, and the grayscale value is set to 0 to obtain the cloud clutter background suppression polarization image. Then, using geometric feature constraints, the final aircraft target detection result is obtained by setting a reasonable aspect ratio threshold and the contrast threshold between the target and the background, removing non target areas such as cloud edge clutter and ripples. Two indicators, accuracy and recall, were used to evaluate the detection performance. The experimental results showed that the clutter suppression and target detection methods studied in this paper can effectively suppress the adverse effects of cloud clutter on detecting aircraft targets in infrared polarized images under different thickness cloud sky scenes, and the detection effect is significantly improved.

Due to the problems of color distortion and low contrast in the acquired underwater images, an underwater image enhancement algorithm based on unsupervised adaptive uncertainty distribution is proposed in this paper. Firstly, the fog removal module and three traditional enhancement algorithms are used to generate different underwater reference images. Then, a multi-color space stretching module guided by statistics is used to correct the color of the reference image and enhance the contrast of the image. Then the image detail features are obtained by the feature extractor. Finally, Progressive Adaptive Instance Normalization and Laplacian convolution are combined to enhance statistical features, enhance image edge texture and improve image visual effect. Experimental results show that the proposed method can effectively improve image contrast and color deviation, and compared with the existing algorithm, the peak signal-to-noise ratio index of the proposed algorithm is increased by 20.78%, the structural similarity index is increased by 4.23%, and the underwater color image evaluation index is increased by 10.06%.

Infrared imaging technology is widely used in military and civilian fields, but in practical applications, accurate and effective detection and tracking of infrared small targets is a bottleneck problem that needs to be solved urgently. In response to the problem that traditional algorithms are difficult to handle complex scenes with low signal-to-noise ratio and deep learning algorithms rely heavily on data, the proposed algorithm combines traditional algorithms with deep learning algorithms and is applied to detect and track infrared moving targets in various complex scenes, with resolutions ranging from 640 * 512 to 320 * 256 video sequences. At the same time, traditional algorithms include both single frame and multi frame detection methods. In order to avoid the problem of poor real-time performance, we selected the TMS320C6678 hardware platform and implemented simulation applications using a DSP+FPGA architecture. Experimental results have shown that this algorithm has excellent performance in object detection and tracking.

Laser absorption spectroscopy (LAS) technology, as a non-contact measurement method with high measurement sensitivity and fast response, combined with tomography technology, can simultaneously achieve visual measurement of flame temperature and components concentrations distribution, and further analyze combustion reaction rules. In this paper, we developed a compact LAS tomography system for the distribution images of temperature and species concentrations. The system mainly consists of two parts, a 32-path sensor and a self-made compact case. In the sensor, 32 pair of collimators with optical fibers are used. The collimator-to- collimator design can help reduce background interference from infrared radiation due to fuel combustion. In the compact case, two distributed feedback lasers, two laser controller units, a 2/1 fiber combiner, a 1/32 fiber splitter and a 32-channels data acquisition unit are integrated. The output lights of lasers are divided into 32 beams by the splitter. The sampling rate of each channel of data acquisition unit can reach up to 1MHz. The measured data is transmitted to the computer for processing and can realize uninterrupted long-term monitoring. In the experiment, a flat flame burner is chosen as the target and the wavenumbers of lasers are scanning around 7185.6cm^-1 and 7444.3cm^-1 respectively. The temperature and water vapor concentration distribution images are reconstructed by the system. The reconstructed image trends of symmetrically distributed. Compared with the temperature testing result of thermocouple thermometer, the temperature distribution trends of the two methods are consistent.

Gas passive infrared imaging technology with high dynamic range (HDR) offers numerous advantages in monitoring harmful gases, including high sensitivity, large detection range, ability to detect multiple hazardous gas types, and strong long-range detection capabilities. However, HDR images cannot be directly displayed on standard devices and need to be compressed to an 8-bit data width for visualization. Improper compression may lead to loss of detail and reduced contrast. Additionally, gas passive infrared imaging often suffers from poor detail, blurred edges, low signal-to-noise ratio, and limited contrast with the background, resulting in overall visual blurring. To address these challenges, this paper proposes an adaptive enhancement and dynamic compression method specifically designed for 14-bit gas infrared images. This method effectively compresses the background information while preserving and enhancing the texture details of weak gas cloud targets. The proposed methodology involves a hierarchical decomposition of the input image into detail and background layers. Subsequently, each layer undergoes separate processing: the detail layer is subjected to enhancement, while the background layer undergoes dynamic range compression. Finally, the enhanced detail layer and compressed background layer are fused to produce an 8-bit image. A comprehensive evaluation is conducted comparing the proposed algorithm with histogram equalization using both subjective visual assessment and objective evaluation metrics. The results demonstrate that the proposed algorithm successfully enhances the clarity of gas cloud targets, improves overall contrast, and effectively suppresses halo artifacts and image noise.

Optical elements are widely used in large optical systems such as laser inertial confinement fusion and astronomical telescopes, but subsurface damage (SSD) generated in the processing of optical components can cause degradation of their performance such as laser damage threshold and imaging quality. Accurate detection of SSDs is necessary to guide the processing of optical components and improve their performance. NDT technology for SSD inspection of optical components based on quantum dots (QDs) tagging breaks through the limitation of traditional NDT process requiring low roughness of optical surfaces and has a broad development prospect in the field of SSD inspection of optical components. Efficient tagging of SSD with QDs can be a key factor to improve the accuracy of SSD detections, but there is a lack of research on how to efficiently tag SSDs with QDs. To solve the above problems, this paper proposes a nondestructive detection technique for SSD of optical components based on high efficiency tagging of CQDs. Firstly, we analyze the characteristics of optical component processing and quantum dot-based detection technology, and determine the factors affecting the effect of QDs on the labeling of optical component SSDs. The effects of different factors on QDs-labeled SSDs were analyzed according to the distribution and fluorescence intensity of SSDs. The results showed that: (1) the changes of pressure and temperature during the process did not have significant effects on the fluorescence properties of CQDs; (2) the excitation wavelength, fluorescence emission peak, solution concentration, solution dose, excitation light power and the pH of the environment in which CQDs are located have certain effects on their fluorescence properties and SSD tagging results. This study provides a useful reference for the efficient labeling and accurate detection of optical components SSD.

In current image fusion techniques, image fusion is usually performed on a dual-band image to obtain a fused image with significant target information, or on an intensity image and a polarization image to obtain an image with stronger visual perception. If more information is to be obtained in a single image, tri-band fusion and intensity/polarization image fusion techniques can be combined. In order to solve the above problems, in this paper, we have acquired some tri-band polarization images through a common aperture multispectral polarization photoelectric device, which contains intensity and polarization visible (VIS) images, as well as the intensity images of near-infrared (NIR) and long-wave infrared (LIR). Besides, in order to obtain good image fusion results, we built an end-to-end self-supervised image fusion network and designed an efficient loss function to train the network. We conducted experiments on TPFNet on the acquired dataset and compared it with other image fusion algorithms. The results show that TPFNet achieves excellent results in both subjective and objective evaluations.

At present, the two-dimensional images obtained by existing small lenses cannot effectively reflect the three-dimensional structure, and many surface defects in4 industry have obvious three-dimensional morphological changes[1]. Therefore, three-dimensional small optical systems are the development direction of industrial optical technology[2]. In view of the characteristics of small lens, small aperture, and single imaging channel, this paper uses a monocular stereoscopic vision method to iteratively reconstruct depth information, and designs a stray light elimination optical system that can obtain both traditional images and three-dimensional stereoscopic images. , improved the photometric stereo vision algorithm used in traditional telecentric systems, developed close-range 3D imaging suitable for small lens systems and improved the accuracy of 3D imaging.

As a commercial aerospace education satellite, “Ladybeetle-I” has flashed many times around the world, which has attracted the attention of many astronomers. This manuscript presents the design of the beacon lamp of this satellite. Consider the limitation of low power consumption, small volume and light weight of “Ladybeetle-I”, this beacon lamp is designed by the combination of light emitting diode (LED) array structure and total internal reflection (TIR) lens. Firstly, through the analysis of structure parameters, color temperature and working state of LED, the light source of the beacon lamp is realized. Secondly, through the analysis of the requirements parameter of TIR lens, the lens of the beacon lamp is realized. This beacon lamp realizes that the “Ladybeetle-I” in the 547-km LEO can be directly observed by naked eyes within the radius of more than 25 km. Furthermore, the satellite in-orbit test results show that the brightness of the beacon lamp at Sub-Satellite Point is higher than -0.6 magnitude (m) stars. To the best of our knowledge, “Ladybeetle-I”, which utilizes this beacon lamp, is the first and only LEO satellite that can be directly observed by naked eye.

Image matching is one of the key technologies of UAV video surveillance. In order to improve the accuracy and efficiency of image matching with different resolutions and scales, a UAV image fast matching method is proposed. Based on the fast SIFT algorithm, the significant areas of the entire image are extracted from the whole image through histogram calculation, and the feature points are calculated on the significant areas. The feature points in two frames are matched, and the mismatched points are eliminated by parallel line algorithm to improve the SIFT feature matching efficiency. The experimental results show that the proposed method not only reduces the matching time, but also improves the accuracy of image matching while maintaining the image matching rate and the robustness of the algorithm, which proves the effectiveness of image matching.

Under daylight conditions, infrared star maps taken by all-day astronomical measurement systems are characterized by high background noise and low signal-to-noise (SNR) ratios. Notably, the accuracy of infrared star point centroid extraction directly affects astronomical measurement system indicators. Traditional infrared star map processing methods focus on SNR enhancement and star point detection but do not fully consider star point irregularities. Notably, SNR ratio enhancements easily alter star point energy distributions, seriously limiting the accuracy of centroid extraction. This study provides a novel method for meeting the high-precision requirements of centroid star point extraction from all-day astronomical measurement systems. First, we extract suspected star point boundaries using the background prediction method. Then, we expand the suspected boundaries using an expansion algorithm for improved accuracy. Finally, we superimpose the multi-frame star map extracted by the expanded boundaries to avoid the SNR increase inherent in traditional superimposed infrared star maps. The superimposition also solves the problem of irregular single-frame star point imaging. Experiments show that the extraction accuracy of our method is 69.2, 73.9, and 33.3% higher than contemporary adaptive threshold segmentation, background prediction, and multi-frame superposition methods, respectively. There are also clear improvements in the success rate of star point extraction.

Hydroelectric dams have a vital role to play in renewable energy and water resources, and ensuring the safety and structural integrity of these dams is crucial. Cracks are a major disease hazard threatening the safety of dams, usually originating on the surface of the dam and continuously extending inwards under the action of hydraulic splitting. Timely and accurately detecting and identifying dam surface cracks is important for maintaining dam safety. However, due to the complexity of the underwater environment and the diversity of dam surface cracks, manual observation methods are time-consuming and laborious to classify underwater surface crack images of dams, and the existing methods have low recognition accuracy when used for the crack classification task. For this reason, this paper proposes an algorithm for classifying underwater surface cracks in hydroelectric dams, which classifies the cracks extracted from the image segmentation model of underwater cracks in hydroelectric dams with the help of an improved RepVGG network. A CBAM attention mechanism module is inserted into the RepVGG feature extraction network to extract a more effective feature representation. Due to the imbalance problem of data samples, the class balancing loss function of IB Loss is introduced to achieve class balancing. The experimental results show that for different kinds of cracks, the accuracy of the constructed algorithms is higher than the classical image classification algorithms, and the average classification accuracy of the network is improved by 2.06% compared to the original RepVGG algorithm, which is improved compared to all other traditional classification networks.

Study on the radiation characteristics of infrared jamming target and the anti-interference capability of infrared imaging system, through the analysis of radiation theory, the experimental study analyzes the change of radiation intensity over time, the radiation intensity of different projection directions and different delivery amount over time, the spectral radiation characteristics of the jamming target and the radiation area of the interference target over time. Combined with the development characteristics of infrared imaging system, it provides the anti-interference effective response strategy of infrared imaging system under practical application condition. The results can provide experimental support for enhancing the anti-interference capability evaluation of the infrared imaging system.

At present, deep learning and other artificial intelligence technologies are constantly applied in the field of remote sensing, and deep learning needs massive data for model driving. However, the development of ship datasets based on optical remote sensing images is slow and not very professional. To solve this problem, we propose three stages of ship intelligence interpretation, and construct a new ship data set for the development of the first stage. The data set reclassifies the existing ships, combined with the purpose of ship target identification and the current development direction of military ships, and is divided into six types of labels, including aircraft carriers, escort ships, submarines, dock landing ships, supply ships and other auxiliary ships. The 909 samples were manually annotated, with the training set contains 795 samples and the validation set contains 114 samples. Compared with the commonly used optical remote sensing ship data sets in China, the data set we proposed has following properties: 1. Professional. The data set targets military ships, The data sources from all ships in the world's military ports; 2. Systemic. Three stages of automatic identification are proposed; 3. Classification and rationality. The classification of this data set combines the general ship classification criterias, practical identification needs and the developments of the ship itself. The world's military ships are divided into six categories; 4. Uneven nature. The number of sliced samples of each type of ships is unbalanced; 5. Developability. This data set can consistently increase the number of samples. New classes of ships will emerge in the short terms. The class can also be divided into six categories according to their missions; 6. The openness. All data for this data set were obtained from Internet public data. Through the verification and evaluation of this data set using five target detection algorithms of YOLO series, the accuracy of the target detection model is about 0.67 and the overall accuracy is about 0.76, which verifies the feasibility of this data set for the verification of the ship target identification algorithms.

In this paper, we propose a template matching algorithm that is robust to deformations and background clutters. A weighted assembled similarity measure is constructed to discover the similarity between two different distributions, and a two-step nearest neighbor searching algorithm is designed to provide the feature points with different weights, which makes it more distinctive when calculating the similarity between the candidate image and the template. A local feature descriptor named Progressive Gradient Descriptor is also put forward to encode the input image to a high-dimensional feature map. Experiments on real-scene data prove that the proposed algorithm is competitive in terms of matching accuracy.

As a new detection method,infrared polarization has outstanding advantages such as anti-masking and anti-stealth. In the future, it has a large application prospect on new lead and photoelectric pods. Therefore, studying the infrared polarization characteristics of the typical background in missile range is of great significance for target test and evaluation, target and environmental characteristics, missile test, camouflage protection, etc.In this paper, the principle and method of infrared polarization imaging was first studied,infrared polarization imaging tests were carried out for the typical background in missile range,and the change law of Gobi background polarization over time in a day was obtained.Through experiments, the effectiveness of three polarization enhancement methods, such as polarization, polarization difference and pseudo-color, were verified to improve image contrast, providing technical support for the testing of infrared polarization imaging equipment in the future.

In the actual engineering applications, the loitering missile seeker has a larger velocity to height ratio (V/H), the sensor imaging scene changes quickly, and the flight environment is more complex, which poses a challenge to the end of the target tracking and guidance. Starting from the tracker algorithm itself and the tracking strategy of the seeker system to improve the performance of the tracking function in engineering applications. Based on the traditional Kernelized Correlation Filter (KCF), the image pyramid pooling idea is used to extract multi-dimensional data features, and the scale estimation is added to the KCF tracking algorithm.so as to improve its robustness to rotation and scale.Improve the tracking timing from the perspective of system management strategy, and optimize human-machine interaction. In order to verify the effectiveness of the method, this paper verifies the tracking performance of the seeker in both laboratory and outfield environments, comparing the tracking results with the traditional KCF on three sets of partial frames with obvious scale transformations, and evaluating the Overlap Precision(OP) and the Center Location Error (CLE), the CLE was reduced by 9.5% and the OP was increased by 34%. In the outfield flight, the tracking of moving and static targets in multi-scene and multi-morphology by the missile-mounted seeker was fully verified, which fully verified the change of target scale and the tracking reliability and accuracy under the extreme environment in the outfield.

In hazy weather, the obtained images of optical instruments are severely degraded due to the multiple atmospheric light scattering, which will significantly influence subsequent image processing such as target recognition and location. In this paper, we propose an efficient image dehazing network based on the framework of the Wasserstein generative adversarial network (WGAN). Inspired by the classic U-net network architecture, we first use a transformer-based image restoration architecture Uformer to modify the generator of WGAN. Then for loss function design, according to the requirement of the image dehazing task, the overall network training is constrained from two aspects, i.e., the pixel loss and the adversarial loss. Finally, the synthetic haze dataset was used to train and evaluate the effectiveness of the network. The results show that the proposed method can obtain high quality restored images, which is comparable to some current methods.

Hyperspectral image target recognition, as an important component of computer vision, has broad application prospects in the military field. We conducted research on hyperspectral image target recognition technology based on deep learning, with typical military targets as the main recognition objects. A target recognition algorithm suitable for hyperspectral images is proposed based on the YOLO network model. To address the issue of the YOLO algorithm not being sensitive enough to small targets, the calculation formula for the loss function in the algorithm has been adjusted, and a multi-scale training method has been added. In response to the phenomenon of missed detections when the YOLO algorithm faces multi-objective clusters, five anchor boxes are added to each grid. In addition, the network convergence speed is accelerated by adding a BN layer; Introducing residual networks effectively prevents network degradation; Adding 1*1 convolutional layer can better extract image detail features and improve network recognition accuracy. The hyperspectral image target dataset is obtained by using hyperspectral imaging instruments to shoot multiple targets from multiple angles and directions under different lighting conditions, and labeling them. Divide the data into training and testing sets for training. The trained model was used for recognition experiments on vehicles in hyperspectral images. The experimental results showed that the model met the real-time requirements for target recognition in hyperspectral images, and performed well in the recognition performance of small targets at long distances.

Air target recognition in real-world scenarios has become an important part of the military offensive and defensive systems of various countries. By identifying aerial targets captured by image acquisition equipment and utilizing the obtained information to achieve effective identification of friend or foe, identifying enemy sources, combat capabilities, and intentions, important references are provided for tactical decision-making. With the continuous development of military technology, traditional manual based recognition methods are no longer capable of identifying aerial targets. The role of deep learning related algorithms, which gradually replace traditional image processing algorithms, in the field of target recognition is becoming increasingly prominent. This article will use deep learning methods to study aerial target recognition. Provide a detailed description of the design and implementation process of the Faster R-CNN training model, including model structure, network layers, and feature selection. This model mainly includes a Region Proposal Network (RPN) and a set of convolutional neural networks for extracting target features and predicting target bounding boxes. Build a basic environment for training models, and establish a dataset of air targets related to military equipment. The dataset samples are annotated according to the VOC2007 dataset format. Train the model using a dataset, conduct testing and detection after training, and finally analyze the calculation results of Eval evaluation indicators to further optimize the algorithm model and improve recognition rate. The experiment verified the effectiveness and feasibility of the Faster R-CNN training model, and applied it to aerial target recognition tasks. The experimental results show that the model can quickly and accurately identify aerial targets.

The Optical system has the advantages of high resolution and clear texture, but it has the disadvantages of being unable to penetrate clouds and fog, and being unable to all-weather imaging. The radar has the advantages of ranging, velocity measurement, all-day time, all- weather, etc., but it has the advantages of obvious spots, unclear texture and low resolution. The technology combines the optical with radar organically, learns from each other, realizes composite imaging, and can take into account the requirements of high resolution, ranging, velocity measurement and all-sky time. The design of optical and radar common aperture to realize integrated composite detection has the advantages of the sane imaging time, the same observation conditions, and real-time composite monitoring of moving targets. In addition, it also has the advantages of high accuracy of dual-channel matching association, dual-channel high-precision pointing association, high-precision fusion, high-precision three-dimensional information acquisition, and small volume and weight, and can also meet the needs of fine (resolution and positioning accuracy) intelligence perception under all-day and all-weather conditions in key areas, significantly improving the comprehensive detection efficiency.

This paper presents a zooming and shaping system for a semiconductor laser beam based on aspheric lens group. Firstly, a theoretical model for transforming Gaussian beam into flat-top beam is established from Gaussian beam model and flattop beam model. Secondly, using the model and the derived mapping function formula, the shaping system is automatically optimized with optical design software to realize the conversion from Gaussian beam to flat-top beam. Finally, a zooming optical path with aspheric lens group is designed by optical design software to achieve laser beam zooming and shaping, and a surface type equation with machinability is obtained. Research results: The designed optical system forms a relatively uniform spot in far field, such as 500m, and the divergence angle can achieve continuous change of 8-32mrad. The research results have certain reference value and guiding significance for the shaping and homogenization design of semiconductor laser beam.

Natural camouflage is a crucial defense mechanism for species to ensure their survival. The most common strategies used in natural camouflage are background matching, environmental texture mimicry, and disruptive coloration. The Camouflage Object Segmentation (COS) task focuses on visually segmenting camouflaged objects that blend seamlessly with the surrounding background. This task poses a greater challenge than the Salient Object Segmentation (SOS) task. Currently, most disguised target segmentation models achieve their tasks by enhancing the texture and contour features of the target, fusing various hierarchical image features, or incorporating frequency domain information of the target. However, these existing models are built on a single framework that uses image context aggregation strategies to segment disguised objects. This approach overlooks the rich and diverse texture features in both spatial and frequency domains for targets. To address this issue, we introduce the cue of frequency components information as an auxiliary enhancement method for images. We also design a dual-stream encoder to process and fuse texture features from both RGB images and frequency perspective images, in order to refine the multi-level features of disguised target textures and contour segmentation. Specifically, we propose a frequency-aware aggregation module that fuses multi-scale features of target textures from a frequency perspective, which includes three different scales of offline discrete cosine transform modules and an image fusion module. A dual-stream encoder is designed to explore global and local texture expressions of targets in both spatial and frequency domains, that multi-layer in spatial domain encoder obtains frequency domain features as masks from the corresponding layers of frequency domain encoder. Additionally, a learnable transpose convolution decoder is used to enhance contour capture capability. Experimental results demonstrate that our model achieves state-of-the-art segmentation accuracy.

As a temporary substrate for preparing the Al₂O₃ ion barrier film, the continuity of the organic membrane significantly affects the compactness of the ion barrier film, while the continuity of the organic membrane prepared by water surface spreading method is directly affected by the non-solvent component. In this paper, the effect of ultrapure water, non-deionized water, NaCl solution, NaOH solution and HCl solution as non-solvent component on the compactness of the organic membrane and ion barrier film are studied. The results show that NaCl solution is the best non-solvent component for the preparation the most continuous organic membrane and the most compact Al₂O₃ ion barrier film.

Submunition swarm combat is a trend in modern battlefields. It aims to achieve precise and organized destruction of time-sensitive and mobile target groups in large operational depth, especially in GNSS-denied environments. This approach relies on assessing the target identification, positioning, and real-time damage assessment of submunition. After collecting damaged images, it is necessary to carry out damage information detection, including explosion flames, smoke, and other information, to determine the impact point of submunition and the process of damaging the target. However, when applying these methods to evaluate submunition, the performance of convolutional neural networks to extract target features still needs further improvement. This paper addresses the problem of false changes in images caused by projectile disturbances in complex backgrounds and the low accuracy of damage feature detection. Building upon the CosNet attention neural network, this paper uses an attention mechanism and proposes a damage feature extraction method based on spatiotemporal attention neural networks. This method achieves high-precision semantic segmentation of damage regions in continuous video sequences, providing a foundation for determining the impact point of submunition and assessing the damage effect. Through our simulation and experiment carried out by rocket sleds, the evaluation of submunition in orbital regions achieved realtime target identification and real-time extraction of the flare region, which validated the effectiveness of the spatiotemporal attention neural network in extracting damage regions in actual dynamic environments. This research provides a critical foundation for damage assessment, offering solutions that enhance the accuracy and reliability of realtime change detection in damage regions within high-dynamic environments.

Aiming at the problem that it is difficult to accurately locate the underwater circular pipe orifice for the underwater pipe orifice inspection robot, a visual position measurement method for the underwater circular pipe orifice based on the spatial elliptic cone model is designed. Firstly, a simple and easy solution is proposed for the influence of underwater light refraction on camera calibration, and then a set of image processing procedures is proposed for the problems of low recognition rate and poor image quality of underwater orifice images, which can stably extract the edge contour of the underwater orifice. The algorithm process mainly includes the use of bilinear interpolation to narrow the edge transition region, the use of dark channel a priori defogging algorithm to eliminate the image blurring caused by the water body, and the Canny algorithm combined with elliptic filtering to finally get the target orifice contour. Finally, the spatial elliptic cone model is used to calculate the position and attitude of the target orifice in the camera coordinate system. In order to verify the feasibility of this underwater nozzle position measurement method, a positional accuracy measurement device is constructed using a high-precision rotary stage and a translation stage. The experimental results show that the visual localization method designed in this paper has high recognition and localization accuracy, and can meet the needs of practical localization.

Monocular visual depth measurement technology has an excellent application prospect in the field of reconnaissance and detection in miniature unmanned systems. In this paper, a monocular depth perception method based on zoom image is studied. At first, the basic principle of monocular image depth perception is analyzed according to the geometric optics theory. On this basis, a depth measurement method that combines chain-code method to calculate target area is constructed: two different focal lengths are used to image the object and two independent sets of functions are listed according to the correspondence between the object distance and the focal length. The image area is used to replace the image point vector, and finally the depth of the object is solved. Then, the design and program compilation of monocular zoom measurement algorithm are realized and the measurement accuracy of the algorithm is preliminarily verified by experiments. Finally, three sets of test experiments are designed with the target distance, focal length ratio and target size as parametric variables. The experimental results show that the measurement accuracy of this algorithm appears a downward trend with the decrease of the target size and the increase of the distance, while the monocular zoom measurement can obtain higher accuracy under the condition of large focal length ratio. The research results can provide technical and theoretical basis for the realization of high-performance monocular depth measurement.

Structured illumination(SI) has a wide range of applications in computational imaging in the form of encoding and recomposing image signals. In this work, we explore the applications and advantages of deep-learning techniques in SI imaging problems. (1) In single-pixel imaging(SPI), where SI encodes the target into a sequence of bucket signals, the recovery suffers from loss of imaging quality due to various noises. We propose an unsupervised deep-learning (UnDL) based anti-noise approach, which outperforms conventional single-pixel imaging methods considerably in reconstructing targets against noise. (2) In blind ghost imaging, we propose two hybrid quantum-classical machine learning algorithms and a physical-inspired patch strategy, leveraging quantum machine learning to restore high-quality images where classical machine learning fails. (3) In structured illumination microscopy(SIM), the illumination encodes high-frequency details into the passband of the objective. Existing optimization-based decoding algorithms are sensitive to noise, while learning-based methods have faithfulness issues. We propose a physics-informed deep learning approach, where the re-parameterized network outperforms its counterparts in terms of noise robustness and recovery faithfulness. In general, the proposed deep-learning-based algorithms solve the inverse SI decoding problem by leveraging inherent priors of neural networks. The proposed framework and idea have the potential to be applied in other scenarios with SI and other computational imaging problems.

The widespread utilization of autonomous underwater vehicles in marine science and engineering has underscored the paramount importance of precise underwater navigation and docking capabilities for underwater robots. Underwater vision navigation systems serve as a pivotal technology, encompassing various domains such as computer vision, image processing, and machine learning. Nevertheless, they confront a series of challenges. This review aims to consolidate the most recent research advancements in underwater vision navigation systems, with a particular focus on critical aspects such as beacon design, monocular and binocular vision, and detection algorithms. At present, underwater beacons can be categorized into active and passive types, with the flexibility to choose between them contingent upon the specific task and environment. On the other hand, relative to the prevalent use of monocular vision, binocular vision systems offer the potential to furnish more accurate depth information, and the seamless integration of monocular and binocular vision systems enhances the precision and stability of navigation tasks. Concerning detection algorithms, traditional feature extraction methods and deep learning approaches each exhibit their merits and drawbacks, but deep learning methods demonstrate immense potential in complex underwater environments. By summarizing and analyzing the above research results, the application of each key technology of underwater vision navigation and docking is sorted out, and its further development direction is foreseen.

Quantitative analysis of spray droplet fields plays a pivotal role in various domains, encompassing internal combustion engine combustion diagnostics, equipment spray coating and corrosion prevention, and unmanned aerial vehicle-based agricultural pesticide dispersion. Precise measurement of the spatial distribution of spray droplet fields facilitates accurate control and orientation of spraying, thereby propelling the intelligent evolution of both industrial and agricultural sectors. In light of the substantial dimensions of spray fields, achieving focused imaging of all droplets on the camera imaging plane during reconstruction proves unattainable. Addressing this challenge, this study suggests employing a four-camera array configuration. According to the characteristics of the defocusing blur of spray droplets, the cameras on the array capture images of the droplets from diverse perspectives. Subsequently, these images are merged through a refocusing process. This method offers accurate extraction of out-of-focus droplet centers. Employing three-dimensional cross-correlation analysis, the motion trajectories of the spray droplet field can be inferred with precision.

Underwater 3D reconstruction has important applications in Marine science, underwater archaeology and resource survey. As a structured light system based on active illumination, fringe projection contour method has been widely used in industrial measurement and manufacturing. However, there are still some shortcomings in the application of turbid water environment, such as the refraction of multi-layer media and the scattering and absorption caused by turbid water, resulting in a large area of information loss or reconstruction failure. This paper aims to study the improvement and application of fringe projection profilometry based on structured light projection technology in turbid water. A calibration scheme suitable for underwater system is proposed to compensate the influence of multi-layer medium refraction on the system. In order to verify the feasibility and effectiveness of this method, a reconstruction experiment was carried out in a pool in muddy water. The reconstruction results are analyzed based on the effective point cloud number, measurement error and Hausdorff distance. The results show that even under high turbidity conditions, the number of reconstructed point clouds has reached millions, and the accuracy can be maintained at the millimeter level, which reflects that the proposed method has potential practical value in the field of high-density and high-precision three-dimensional reconstruction in turbid water environment.

Underwater optical imaging technology is extremely important for a wide range of applications, e.g., the operation and rescue, resources exploration and species protection in complex underwater environment. However, the underwater images always suffer from reduction of intensity, loss of contrast and blurred details due to the scattering and absorption caused by the scattering medium. Therefore, effective research methods are needed to achieve clear underwater imaging. In this paper, we propose an underwater polarimetric image descattering and material identification method based on unpaired multi-scale polarization fusion adversarial generative network. In our method, polarimetric image with different polarization states (I₀, I₄₅, I₉₀) and the degree of polarization (DOP) maps are introduced to the network for the purpose of fully extracting polarization information. Experimental results show that our method can remove the underwater scattering effect effectively. And good restoration effect is achieved under unparied data sets, which breaks the dependence on strictly paired polarimetric images in previous learning methods. Not only complete intensity information is retained, but also more polarization information of the target could be extracted from the descattered image, which is especially beneficial for subsequent target identification. The average values of peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) of the descattered images under different turbidity have been greatly improved comparing with the original images. All in all, excellent performance is achieved in both subjective visual perception and objective evaluation indicators using our method, which provides a new way for underwater polarization imaging.

As the battlefield environment becomes more complex, optical guided weapons will face increasingly complex photoelectric interference environment. This paper analyzes the photoelectric interference and interference mode faced by the optical detection system used in surface-to-air missiles, also gives anti-interference methods and simulation results, aiming at providing reference for anti-interference design of surface-to-air missiles.

With the continuous development of fiber positioning technology, the positioning accuracy of LAMOST (Large Sky Area Multi-Object Fiber Spectroscopic Telescope) needs to be further improved. The open-loop control based on calibration curves can no longer meet the requirements of the next-generation fiber positioning technology. Therefore, a closed-loop control system needs to be adopted, and image measurement technology needs to be introduced to improve accuracy. The key to this mechanism lies in correctly matching the positions of the units in the image. The matching process is divided into two steps: first, calibrating the camera and calculating the matching parameters of the working units in the image based on the positions of the reference unit light spots in the image, and then matching the working units. Due to the overlap of the working areas of different units, we can predict the approximate current position of the work unit, and perform matching at those positions to improve the accuracy of the system matching and reduce the complexity of the matching algorithm. The experimental results indicate that even in cases where there are a significant number of errors in the predicted positions due to various reasons, the mechanism still maintains a high success rate in matching. At least 95.3% of the units exhibit an error of less than 40 micrometers after three steps. Ultimately, this matching mechanism not only has excellent robustness but also has low sensitivity to unit parameters, initial positions, and unexpected situations such as collisions or communication issues.

The inverse synthetic aperture LiDAR (ISAL) system demonstrates excellent performance in various application scenarios, including long-range target imaging and recognition. However, due to the micrometer scale of the ISAL system carrier, compensating for the motion error of high-speed complex motion targets becomes challenging. Linear frequency modulation (LFM) signals are not enough to describe these targets accurately, as they can be better represented by high-order polynomial phase signals. Nevertheless, the presence of second-order and higher phase components can greatly affect the coherent synthesis between pulses. To tackle this issue, we propose a non-searching fast cubic phase signal estimation algorithm based on the fourth-order instantaneous autocorrelation function. This algorithm aims to estimate the second-order and third-order coefficients of the rotational error signals using the non-uniform Fourier transform. By accounting for advection compensation, we model the phase of the signal of the complex motion target as the third-order form of the slow time. The Fourier transform of the fourth-order instantaneous autocorrelation function presents challenges, such as the nonlinear coupling between the slow time and delay time, as well as the nonuniform sampling of the delay time. We suggest utilizing the nonuniform Fourier transform to complete the two-dimensional transformation and estimate the obtained second-order and third-order coefficients. In comparison to the traditional range doppler (RD) algorithm, our proposed algorithm achieves better image point focusing and enhances the energy aggregation of scattering points by approximately four times. Simulation results validate the effectiveness of this algorithm.

Factors such as scattering and absorption of light by suspended particles and lack of light in deep water exist in complex underwater environments, leading to visual degradation effects such as loss of underwater image features, colour deviation and contrast reduction. With the development of artificial intelligence, deep neural networks are widely used in the field of computer vision and show their powerful brain-like separation (local information processing) and integration (global information processing) processing capabilities. In this paper, we use the visual saliency model to construct a Gaussian pyramid of luminance, orientation, edge and colour applicable to underwater degraded images to obtain shallow image features of underwater images. Combined with the VGG16 convolutional neural network model to construct a progressive enhancement neural network based on deep learning, which in turn improves the high-dimensional saliency features of underwater degraded images. The experimental results show that the enhanced underwater image features of this algorithm have better detail retention and the colour is more in line with the human eye vision, and the experimental results of the objective indexes are better than the comparison algorithm.

A novel optical scene simulation method is proposed in this paper which is designed for evaluating intelligent driving systems incorporating large-array flash lidars as the 3D detector. A 3D scene is first decomposed into discrete time slice patterns and displayed on a spatial light modulator. An integrated imaging system is then used to reconstruct all time slice patterns in space into a 3D scene. We have developed a 3D lidar scene simulator with a spatial resolution of 640×512 and a distance simulation range from 0.5 meters to 32 meters, achieving a distance simulation resolution of 15mm. The method proposed in this article can significantly reduce the time and space requirements for the end-of-line testing of lidar products on automated production lines and provides an effective approach for comprehensive performance evaluation in the field of intelligent driving systems.

The infrared digital Fourier transform spectrometer has several advantages, including small size, light weight, high stability, increased throughput, and enhanced spectral resolution, making it a valuable tool in the biomedical field. The data acquired by this instrument directly is interference data, and the required spectral data is obtained through a spectral recovery algorithm. The reconstruct spectral is determined by the acquired data quality and spectral recovery algorithm. However, the type of silicon photonics-based Fourier transform infrared spectrometers often encounter non-uniform optical path differences in collected interference data, due to the limitations in hardware design and manufacturing processes, leading to the spectral obtained by commonly used spectrum reconstruction methods inaccuracies. In this paper, a spectral recovery algorithm based on deep neural networks is proposed for the reconstruction of spectra from non-uniformly sampled interference data, Compared to other spectral recovery methods, the proposed method achieves better spectral angle (SA) and relative quality error (RQE) between the reconstructed spectra and ideal spectra.

Detecting oil leakage faults in traction transformers is critical to Electrical Multiple Unit (EMU) operation fault detection. Mineral transformer oil, which exhibits fluorescence characteristics, is commonly employed in EMU traction transformers. The position and intensity of the fluorescence characteristic peaks determine the selection of excitation light source wavelength and imaging spectral band for the mineral transformer oil leakage detection technology based on fluorescence imaging. However, the fluorescence characteristics will change with ageing. To investigate the changes in the fluorescence characteristics of mineral transformer oil used in EMU during the ageing process, accelerated thermal ageing experiments were conducted in the laboratory. The three-dimensional fluorescence spectra of samples were collected regularly. Furthermore, the hydrocarbon group composition and interfacial tension of the samples were routinely detected, with the intent of unraveling the underlying factors contributing to the variations observed in the fluorescence spectra. The results show that the position of the fluorescence spectrum peak gradually shifts to the long-wave direction with the prolongation of the ageing time, and the intensity of the peak continuously decreases. The experimental results offer the experimental and data foundation for the selection of excitation light source wavelength and imaging spectral band in the application of fluorescence imaging detection technology to detect oil leakage faults in mineral transformers of EMU.

This research conducts a comprehensive study of a laser coherent detection system utilizing balanced detector. The investigation explores the effects of key factors on signal-to-noise ratio (SNR), such as the fiber optic coupler’s coupling ratio (CR), quantum efficiency matching ratio between the diodes of the balanced detector (QER), the local oscillator (LO) optical power. A self-built 1550nm vibration measurement laser system was built to detect the indoor acoustic vibration. Conduct time-frequency analysis of the vibration signal, and evaluate the results through the evaluation indicators, sliding window spectral estimation (SWSE) to calculate the SNR. The results show that when the LO optical power approximates 1.47mW, the SNR are in the optimal range. This research not only enhances the expansion of later target classification to fine recognition, but also provides reference value for various applications.

Chemical component analysis is an indispensable part of steel analysis. At present, the remote intelligent detection of the component analysis of steel needs to be further developed. In this study, laser‐induced breakdown spectroscopy (LIBS), a sensitive optical technique that can rapidly analyze various elements, is applied to real‐time detection and steel analysis of steel at elevated temperature environment of a steel industry. Representative steel samples are selected for spectral analysis. The characteristic spectral lines of different elements observed in the spectra. By comparing the spectra of different samples, the differences among them discussed. First, principal component analysis (PCA) is used for dimensionality reduction. With the help of PCA, the data are distributed in two dimensional and three-dimensional space and different kinds of the constituent elements of steel are classified successfully. Then, combined with error back propagation training artificial neural network, the steel identification model was established, and the recognition rate can be improved. The results show that LIBS is a powerful tool for remote intelligent real-time detection and classification, and has great application prospects in the chemical component analysis of steel at elevated temperature environment in the future.

Human physiological parameters such as heart rate (HR) and heart rate variability (HRV) hold significant importance in health analysis and disease diagnosis. Traditional measurement methods heavily depend on contact sensors, which are inconvenient and uncomfortable for subjects. In contrast, video-based remote photoplethysmography (rPPG) has emerged as a promising technology for non-contact physiological measurement. However, its utility is compromised by susceptibility to motion artifacts and lighting. In response to the challenges posed by motion artifacts and lighting, we propose a robust rPPG framework. This paper assesses the noise robustness of diverse methods employed in rPPG-based physiological measurement, including face detection and tracking, region of interest (ROI) selection, rPPG signal extraction and heart rate calculation. The experimental results demonstrate the efficacy of the proposed framework in enhancing pre-processing outcomes and facilitating accurate heart rate measurements, even during dynamic motion tasks.

The three-dimensional particle image velocimetry (3D PIV) technique, as a non-intrusive method for three-dimensional full-field velocity measurement, has garnered extensive utilization across diverse domains including biomimetic dynamics, combustion diagnostics, and the structural design of aerospace equipment. Synthetic Aperture Particle Image Velocimetry (SAPIV), based on camera arrays, digitally merges images obtained from different perspectives to simulate the imaging effects of large-aperture cameras. This approach allows for large-scale, high-resolution flow field measurements. In this study, the three-dimensional intensity characteristics of particles within sequences of refocused images are investigated. Leveraging the spatial distribution patterns of grayscale information for individual focused particles, we designed a three-dimensional convolutional neural network (3DCNN) capable of extracting focused particle positions. Throughout the particle extraction procedure, this three-dimensional CNN network systematically analyzes the sequence of refocused images and subsequently derives both particle positions and grayscale information for focused particles based on their distinctive characteristics. The tracer particle field in simulated experiments were reconstructed and the reconstruction quality was evaluated. The results demonstrate the high precision of our proposed method in reconstructing three-dimensional tracer particle information in SAPIV.

In the three-dimensional(3-D) microscopic measurement based on white light interference imaging, due to the influence of external environment interference and the phase-shifting step size of micro-shifter, the sampling signal by longitudinal scanning may have the problem of "peak leakage", which seriously affects the accuracy of measurement results. In this paper, by analyzing the spectrum of interference signal, a method for extracting zero optical path difference(OPD) position based on multinomial Gaussian fitting for envelope curve of filtered time domain signal is proposed. Finally, based on the method proposed in this paper and ZYGO commercial interferometer, the comparative measurement experiment of standard steps is carried out. The simulation and experimental results show that the method has good robustness and effectiveness.

Radiative heat transfer is a significant heat transfer mechanism in the formation of engine exhaust plumes, resulting in varied effects on the exhaust plume flow field and infrared radiation characteristics. In this work, simulation model is used to simulate exhaust plume flow of solid-propellant rocket motor, with consideration given to the recombination effects of 9 species and 10 elementary reactions. The study investigates the impact of radiative heat transfer on the exhaust plume flow field at different altitudes. Furthermore, a narrow-band model and the Line of Sight (LOS) approach are employed to calculate the infrared radiation characteristics of the flow field. It is indicated that radiative heat transfer affects the temperature of the flow field and the distribution of intermediate products such as CO and OH. Notably, at flight altitudes of 10km, the impact of radiative heat transfer is relatively weak, while at an altitude of 60km, the effect is more pronounced, which is attributed to the rarified air conditions prevalent in high-altitude regions, where molecular collisions are less frequent, thereby radiative heat transfer plays a significant role in heat conduction. Furthermore, radiative heat transfer also exerts a noticeable impact on the infrared characteristics of the flow field.

With the increasingly widespread and important application of optics in modern human life and production, research on the propagation characteristics and correction methods of light beams in the atmosphere near the ground surface is becoming increasingly important. The complex atmospheric environment near the ground can affect the propagation of light beams, causing effects such as energy attenuation, beam expansion, beam drift, intensity flicker, phase fluctuations, etc., reducing the quality of beam propagation and affecting the performance of optical applications in various fields. A review was conducted on the impact of near-Earth atmosphere on beam propagation and its correction methods, mainly elaborating on the effects of atmospheric absorption, scattering, and turbulence on beam propagation. Correction methods such as adaptive optics, large aperture receiving, spatial diversity, partially coherent light propagation, and image processing are proposed.

The defects of solar cells (such as crack and finger failures) are commonly detected by electroluminescence (EL) methods combined with image-based artificial intelligence algorithms.In order to improve the defect detection performance of high-resolution EL images, this research introduces a modified Feature Pyramid Network (FPN). The Content Integration Module (CIM) and an Attention-Guided Module (AM) are combined to enhance the balance between the feature representation and the receptive field. A detector is also constructed by incorporating the above module into the traditional FPN, allowing the region feature network to receive multiscale features with enlarged receptive fields. The performance of the proposed detector algorithm is compared with other detectors including FPN, PANet, BAF-Detector. The paper also analyzes the issues associated with the datasets. We discussed the crack as the research object. Common mismarked conditions are identified and corrected. We note the unevenness of the crack geometry in the data set. In various geometric features, the distribution of crack length and its effect on the detection efficiency are discussed. Simulation data sets with different distributions were constructed and numerical experiments were carried out. The simulation data set method can be used to test the performance of the algorithm. The research in this paper will provide useful ideas for other defect detection fields

Underwater in-situ darkfield microscopy imaging can grasp excellent resolution and high contrast images for observing transparent and living plankton. However, its limited field of view and shallow depth of field leads to a large amount of redundant data during image acquisition. Manual extraction of regions of interest (ROIs) imposes significant labor and time costs. Therefore, we develop an automatic ROI extraction method to obtain high-quality in-focus individual plankton images from raw images. A multiscale underwater in-situ microscopy system (MUIMS)based on darkfield imaging has been established. The system allows for the monitoring of plankton in the size range from 10μm to 1cm. A series of preprocessing steps are performed on the raw darkfield images to make edge detection and segmentation of plankton regions. This can realize rapid extraction of ROIs from the raw images. A dataset is constructed from these ROIs, and each image is labeled to indicate whether it is a high-quality image. By using a transfer learning strategy, we improve a pre-trained ResNet-18 model and train it to accurately classify high-quality ROIs. The experimental results show that this method can rapidly extract useful information from raw images, achieving an accuracy of over 90% in judging the quality of ROIs. This research is beneficial for building a dataset of high-quality plankton, which provides a crucial data foundation for ecological research and marine biodiversity conservation.

Optical sparse aperture imaging systems require a very accurate control of the phase errors between sub-apertures to achieve high-resolution. The linearized analytic phase diversity (LAPD) is a fast technique to co-phase sparse aperture imaging systems. However, it requires focal plane and out-of-focus images to achieve phase diversity, so the light path is complex. In this Letter, we propose a spatially modulated analytic phase diversity algorithm. The diversity images are obtained by the spatial modulation diversity technology (SMDT). The validity of the method was verified by simulation experiments. The method can typically detect piston errors within |π/2| rad rms in the extended target scenes. Co-phase accuracy can be achieved by iterating up to λ/60. The method does not require additional splitting device, which is conducive to system miniaturization, while the required calculations are simple and the piston detection is fast. This method can give a simple, fast and low system complexity solution for fine phasing of sparse aperture imaging systems under any imaging scenes.

In order to study the recognition methods of underwater targets in harsh environments and improve the accuracy of image based underwater target recognition, a deep learning based underwater target recognition method in harsh environments was carried out. Propose a simulation method for underwater target harsh environment images based on clearer image samples. The simulation processing mainly includes adding random scale and spatial position obstructions, adding random noise, and obtaining simulated image data. Establish an integrated image dataset, which includes clearer original image data and simulated image data. The recognition results of target recognition models trained on the integrated image dataset have significantly improved compared to the original image dataset.

The imaging saturation can lead to errors in image processing and result in incomplete or inaccurate three-dimensional (3D) outcomes. To address this problem, we propose an adaptive fringe luminance projection method based on iterative luminance adjustments to the pattern regions corresponding to the highlight targets. After 3-4 times iterations to update the intensity distribution of the patterns, we can ensure that the captured images are without overexposed area; thus, the intensity variations provide valid information for the subsequent 3D reconstruction. The experimental results show the accurate and complete 3D reconstruction for pathological sample surfaces, which reaches the requirements for 3D digitalization of pathological samples.

When the large-aperture antenna pose is measured by the laser method, the subreflector has a strong energy due to the convergence of the antenna itself and the instability of the near-ground atmosphere, and then the laser signal is easily drowned out. Atmospheric attenuation and other factors will weaken the laser transmission in different spectral bands, which reduces the recognition accuracy of laser spots. Aiming at the research of measuring the laser spectrum of large-aperture antenna, this paper analyzes the influence of atmospheric attenuation on the signals of different laser spectrum bands under the fixed distance of subreflector measurement, and compares the solar radiation energy of different wavelengths in laser. Finally, the stray light simulation analysis and experiments are carried out on different laser working spectrum bands to verify the accuracy of the spectrum research used in antenna measurement. Experiments show that the 1064nm wavelength spectrum, as the working spectral band of the antenna measurement, the spot information is more obvious, which can effectively realize the extraction and identification of the spot center and provide a strong guarantee for the antenna pose measurement.

Liquid crystal polymer polarizing gratings have been widely studied in augmented reality optical waveguide displays due to their nearly 100% diffraction efficiency, large angular bandwidth, and wavelength bandwidth. The thickness of each area of the grating film layer prepared by the existing film production method (spin coating) is the same, which is unfavorable for uniform imaging of large-format liquid crystal polymer polarized grating optical waveguides. This paper proposes a method of etching and thinning the grating film layer to achieve specific efficiency modulation. The experimental values and simulated values are in good agreement, providing a reference for segmented control of the efficiency of large-format liquid crystal polymer polarization gratings.

As a new photoelectric detection method, polarization imaging can effectively improve the detection range and recognition accuracy of key targets under harsh environmental conditions by its excellent imaging effect of de-fog reconstruction. This paper proposes an image detection and recognition network based on polarization information and intensity information. The network is based on the yolov5 network integrated with DYHEAD block for recognition, realizing scale perception, space perception and task perception in a unified manner. Meanwhile, Res2Net is integrated for multi-scale characterization at the granularity level. Finally, the attention mechanism is introduced to realize the adaptive extraction and fusion of multi-scale features at the granularity level. The results show that the proposed method can effectively improve the recognition accuracy in low visibility environment.