Research On Image Semantic Segmentation Algorithm Based On Deep Learning

In recent years,image semantic segmentation,as a research hotspot of computer vision,has received extensive attention.Its application scenarios include industrial,medical,military and other fields,involving human-computer interaction,assisted driving,medical image diagnosis and other key technologies.With the popularization of intelligent vehicles,semantic segmentation,as a core technology in the field of autonomous driving,not only needs to complete the recognition and understanding of street scenes,but also needs to accurately analyze pedestrian movements.To address the high cost of manual annotation in image semantic segmentation datasets and the difficulty in accurately recognizing pedestrian movements,this paper proposes a semi supervised image semantic segmentation algorithm based on generative adversarial networks and a human image semantic segmentation algorithm based on graph neural networks.The main work and innovation points are as follows:First,in order to solve the problem of high cost of manual labeling,this thesis proposes a semi-supervised image semantic segmentation algorithm based on generating confrontation network.The network uses both labeled data and a large number of unlabeled data.When using unlabeled data,the probability map generated by the discriminator is binarized by a threshold value,and the intersection is used to highlight the areas with high similarity in the prediction map,which are regarded as false labels of unlabeled data,so as to realize semi-supervision.Secondly,in order to solve the problem of label misjudgment and local pixel misjudgment in the baseline network,a macro and micro discriminator is proposed to achieve both deep semantic information and shallow detail information.By adding a new discriminator network branch,the low-resolution map with richer semantic information in the generator network is discriminated.At the same time,the original discriminator network is improved to enrich the details of false tags.Thirdly,in order to solve the problem that the convolutional neural network cannot effectively extract human body structure information,a node specific network is designed,which is composed of node extraction module,low-high module and multi-set joint module.The node extraction module can specifically extract the features of various parts of the human body as the nodes of the graph convolution network.The low-high module can model the relationship between various parts of the human body at three levels:low,medium and high.The multi-set joint module divides the human body parts into three sets:part set,half-body set and full-body set,and enhances the human body modeling by mining the construction relationship between the sets.Fourth,in order to improve the model expression and adaptive ability,this thesis takes the channel importance as the characteristic basis of the graph node,and directly learns the influence of each channel on different nodes from the backbone network characteristics,so that the node characteristics can self-learning.At the same time,in the low-high module and the multi-set joint module,the dynamic convolution core is generated by aggregating multiple parallel convolution cores to adaptively adjust the convolution parameters according to the input image.Experimental results show that the semi-supervised image semantic segmentation algorithm based on generative adversarial network solves the problem of insufficient dataset to a certain extent.The algorithm improves on label misidentification and target edge detail compared to the baseline.In addition,the effectiveness of modules and the rationality of network hierarchy settings in the network of specific nodes are verified by ablation experiments,and better segmentation results are obtained in the semantic segmentation of human images.