With the development of remote sensing technology, remote sensing images of buildings are of great significance in urban planning, disaster response, and other directions. When we use a neural network containing batch normalization layers for semantic segmentation, the neural network is sensitive to batch size and has low segmentation accuracy for occluded and dense buildings. This paper proposes a method for building segmentation in remote sensing images based on Nested UNet (UNet++) deep neural network. First, the UNet++ network is used to extract features, and the Group Normalization (GN) method is used instead of Batch Normalization (BN) to alleviate the model's sensitivity to batch size. Then, the weighted combination of Cross-Entropy Loss (CELoss) and DiceLoss is used as the loss function to improve the feature extraction ability of the neural network for unbalanced buildings. Finally, experiments are carried out on the WHUBuilding dataset. The experimental results show that the improved model (UNet++-GN) improves Mean Intersection over Union (MIoU) and Mean Pixel Accuracy (Macc) by 12.16% and 2.92%, respectively, compared with the original model (UNet++-BN).
Aiming at the problem of model instability and overfitting of deep neural networks with the deepening of the number of network layers, the current mainstream method is to use batch normalization (BN) to alleviate them. However, since the BN method is more sensitive to batch size when the batch size is small, the model performance will be poor. For a relatively large model, due to the limitation of video memory, the batch size cannot take a large value, limiting the model's performance. Because of the dependence of BN on batch size, this paper adopts group normalization (GN) instead of batch normalization (BN) in the UNet network to alleviate the impact of the model on batch size. Then experiments are carried out on the WHUBuilding dataset. The experimental results show that the improved model (UNet-GN) improves the mean intersection over union (MIoU) and mean pixel accuracy (MPA) by 10.66% and 1.65% respectively compared with the original model (UNet-BN).
To effectively realize the reasonable obstacle avoidance of the detection robot, VGG based obstacle discrimination method is proposed. Above all, the image captured by the robot is input into the multi-layer convolutional neural network to obtain the high-level image features, which are used to construct the more accurate neural network model parameters and to train the softmax classifier with these parameters. Then the distance between the imported images and the data images is calculated by using the softmax classifier, and the similarity between the obstacles and non-obstacles is estimated. The experimental results show that the discrimination accuracy increase to above 94%. And the proposed method is more effectively compared with traditional ultrasonic and radar methods.
In order to solve the problems of unstable training and texture blurring of generated images, we proposed a generative adversarial network combining residual and attention block. The attention module is added to the network, which reduces the dependence on the network depth and reduces the depth of the model. The dense connection in the residual module can extract richer image details. The number of parameters is reduced and the calculation efficiency is greatly improved. Generative adversarial network is used to further improve the texture details of the image. Generator loss functions include a content loss, a perceptual loss, a texture loss and an adversarial loss. The texture loss is used to enhance the matching degree of local information, and the perceptual loss is used to obtain more detailed features by using the feature information before an activation layer. The experimental results show that the peak signal to noise ratio is 32.10 dB, and the structural similarity is 0.92. Compared with bicubic, SRCNN, VDSR and SRGAN, the proposed algorithm improves the texture details of reconstructed images.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.