Research On Graph Convolution Technology And Its Application In Weakly Supervised Semantic Segmentatio

In recent years,the popularity of deep learning has driven the development of semantic segmentation.This technology has widely application prospects in automatic driving,intelligent medical treatment,etc.Compared to other coarse-grained tasks,semantic segmentation requires fine dataset labels.In the process of labeling,it will spend a lot of time and expensive human cost,and there is a certain error rate in manual methods.Weaklysupervised semantic segmentation can complete the task of semantic segmentation only by using simple annotation types(such as image level labels,boxes,graffiti,etc.).But this level of label can only provide the lowest semantic information,so the accuracy of weakly-supervised semantic segmentation is still unsatisfactory.With the proposal of graph convolution,graph neural network entered people’s attention again.Compared with the limitations of convolutional neural networks(the input data is required to be European data),graph structure data is more suitable for the real world.The use of adjacency matrix in graph convolution makes the model global,which also makes up for the disadvantage of insufficient receptive field in convolution operation.Based on the above background,this paper proposes the application of graph convolution technology in weaklysupervised semantic segmentation to solve the dilemma of low accuracy of the current weaklysupervised semantic segmentation model.The specific research work is as follows:Firstly,a weakly-supervised semantic segmentation method based on class-related graph convolution is proposed.In this method,the graph convolution operation is carried out for the class activation map to achieve the purpose of activating region growth.We map the feature vector to the graph space,use the prior knowledge to construct the adjacency matrix,and use the parallel graph convolution method to learn the region extension strategy specific to each class of objects.We try different graph convolution methods to get better segmentation results.Secondly,a weakly-supervised semantic segmentation method based on node classification is proposed.The pixel classification of the image is regarded as the node classification of the graph structure.Performing graph convolution operation on the extracted feature vector,and the generation of class boundary map is used to construct the affinity relationship between pixels.A self-correcting module is proposed to reduce the influence of noise in OAA method and make the class activation map more reliable.The original graph is clustered into different subgraphs by graph clustering,and the graph convolution operation is carried out on the subgraph.Finally,the classification results are returned to the original graph to solve the problem of large amount of graph convolution calculation.Finally,a weakly-supervised semantic segmentation method based on pyramid graph convolution is proposed.We propose a new idea for model optimization.Because the model task is image classification,it will not actively identify more accurate object areas.By artificially applying pressure and using the idea of pyramid,the graph convolution operation is carried out on the feature vectors of different sizes,to obtain a more powerful feature vector representation.In addition,we use the idea of total aggregation to obtain the adjacency matrix by aggregating the height,width and channel information.