Image Semantic Segmentation Based On Generative Adversarial Networks And Self-Attention Mechanism

Semantic segmentation is a visual scene understanding task.Its target is to predict the category label of each pixel in the input image,so as to achieve object segmentation at the pixel level.Semantic segmentation is widely used in autonomous driving,robotics,medical image analysis,and video surveillance.Therefore,it is of great significance to improve the accuracy of image semantic segmentation.It is found that the accuracy of image semantic segmentation is affected by the number of labeled samples and the model structure design.To a certain extent,the current research methods have some problems,such as requiring large amounts labeled data and inadequate use of context information.To deal with above problems,this thesis has carried out the following two aspects of research work.First,an adversarial training method is proposed to enhance the performance of semantic segmentation model.We train the full convolutional semantic segmentation network and the discriminant network simultaneously,and improve the segmentation accuracy by detecting and correcting the high-order inconsistency between the real label and the segmentation graph.The whole semantic segmentation network is trained by coupling segmentation loss and adversarial loss.In order to solve the problem that semantic segmentation data sets need a lot of manual annotation,based on the fact that adversarial learning can enhance the segmentation results,we use the discriminant network to extract the trusted regions in the segmentation results and generate pseudo labels to assist the network training,realizing semi-supervised image semantic segmentation.Second,a semantic segmentation model based on self-attention mechanism is proposed.In order to adaptively extract the context information and capture the spatial structure information which gradually gets lost with the deepening of the network level,the self-attention mechanism is introduced into the image semantic segmentation and combined with the atrous spatial pyramid pooling module.The attention module adaptively extracts the global context information according to the different scale feature objects extracted by the atrous spatial pyramid pooling module.In order to overcome the problem that the classical attention model has too many parameters when extracting the context information,the criss-cross attention method is used,which only extracts the horizontal and vertical information.By concatenating this module,we can capture the environment information from the whole image dependency in a more effective way.