Image Semantic Segmentation Method Based On Graph Convolution

Image semantic segmentation is the most basic understanding task in autonomous driving,robotic navigation and agricultural science.Each pixel in a large high-resolution image is grouped into a set of semantic tags.Over the years,neural networks have been proved as a successful solution,but the current work is mainly focusing on the convolutional network With the deepening of the research,the objects in many scenarios show very large scale changes,which poses great challenges to the global representation and local details:(1)In large-scale images,the background is noisy,and many small objects of great significance to the segmentation task,for instance,the edge information of traffic lights cannot be recognized accurately due to the occlusion of other objects,which is easy to be submerged in the intricate background,resulting in unsatisfactory segmentation results;(2)Convolutional neural network is good at relying on convolution operation to process local information,but it ignores contextual information,lacks global description,which leads to poor performance in obtaining global relations between distant regions,and is limited by the perception field of convolutional network,resulting in incomplete recognition of large objects,rough edges,and semantic confusionAiming at the problem of semantic segmentation exposed in global representation and local details,this paper designs a set of semantic segmentation methods Similar to Inception Multiple parallel Graph convolution(STIMG).The method consists of optimized ResNet network and STIMG module.The optimized ResNet network combines skip structure and carries out channel screening according to two-dimensional entropy;in the STIMG module,the feature map with more semantics and richer channel is obtained by multi-sampling,and the function of GCN is enhanced by kernel function and error function,and the calculation efficiency is optimized by using one-dimensional convolution in the calculation process;finally,the optimized ResNet network is adopted as the main system and the STIMG module is embedded in it.The main innovations of this method are as follows(1)To solve the problem that the semantic segmentation method is difficult to communicate the context information in a large range,STIMG module realizes the global information inference and local details enhancement between arbitrary regions,especially in the image with chaotic background,it improves the segmentation accuracy of inconspicuous objects and large objects,and finally achieves the goal of fine edge segmentation(2)The optimized ResNet network makes the abstract semantic information and the easy-to-understand appearance information complement each other,and provides semantic feature maps with various levels for segmentation tasks and helps us to get more detailed segmentation results(3)STIMG is convenient,can be conducted end-to-end training,is not sensitive to input specifications,can be inserted into the mainstream of the deep network with a little change,improve network performance,especially can improve the inconspicuous and large object segmentation accuracyIn this paper,the Cityscapes data set was used to test STIMG.Under the condition of using a smaller training set,and on the premise that the average result equates with the most accurate semantic segmentation model,the problems of rough recognition of inconspicuous objects and chaotic segmentation of large objects in the existing model were greatly improved,and the goal of refining the segmentation edge was achieved.