Research On Image Semantic Segmentation Method Based On Contextual Information And Weak Annotation

The research on computer vision technology continues to increase,and semantic segmentation belongs to one of the hot research directions at present.Allocating a separate class label to each pixel of an image is an important step in building complex unmanned systems.Although image semantic segmentation based on deep learning is developing rapidly,there are still some shortcomings such as complex network model,large number of parameters and large computation.Moreover,the development and application of semantic segmentation model are constrained by the difficult and expensive pixel level labels.In this paper,the number and computation of semantic segmentation network parameters and the excessive dependence of model training on data pixel level labeling are studied.The main work and innovations are as follows:1.To solve the problem of slow reasoning speed of semantic segmentation network due to the large number of parameters and high computational complexity,this paper proposes a lightweight image semantic segmentation network based on contextual information.By using conditional channel weighted block lightweight network,the number of parameters and computation amount of the model are reduced while maintaining the high resolution feature diagram.By adding contextual information module,we can enhance the feature representation which is favorable to segmentation task,and ensure the accuracy of network segmentation.Experimental results on Cityscapes,ADE20 K,and PASCAL VOC2012 datasets show that the proposed segmentation network achieves a good balance between parameter number,computation,inference speed,and segmentation accuracy,and is superior to mainstream semantic segmentation networks.2.Aiming at the problem that the semantic segmentation model relies too much on pixel level labeling and restricts the application of the semantic segmentation model,based on weak labeling information I propose a semantic segmentation network.Based on the twin structure of SEAM network,the pixel adaptive residuals network is designed to learn specific content information by using pixel adaptive convolution in order to retain spatial information characteristics of the target.Then,a dual correlation module is designed to enhance the ability of the network to learn global information,so that the network can generate a centralized and complete target location map based on image level weak label using class activation map technology,and obtain the semantic label of image data.The segmentation performance is 66.9%and 66.2% experiments on the PASCAL VOC2012 in the training and test sets,respectively,compared to other mainstream methods,and our method generates better semantic labels for image data.