Research On Image Semantic Segmentation Algorithm Based On Deep Convolutional Neural Network

As one of the most important tasks in the field of computer vision,image semantic segmentation has been widely applied in many fields in real life.The meaning of image semantic segmentation is to assign categories one by one to each pixel in the image.However,training convolutional neural network with full supervision requires manual labeling of each pixel,which is time-consuming and laborious.Compared with full supervision,weakly supervised semantic segmentation has the advantages of fast labeling speed and low labor cost,which has attracted the attention of many researchers.This paper is based on image-level label weak supervision.Seed growth and boundary limit algorithm based on simple clear train of thought and good segmentation results two advantages,has been weak supervision level of many scholars in the image tag semantic segmentation algorithm are applied,this paper also based on this method,classification and segmentation of Deep Seed Region Growing(DSRG)model network module is improved optimization respectively,this article mainly working content is as follows:In view of the deficiency of sparse initial seeds in the studied DSRG model,the classification network part of the DSRG model was improved.The improved model is to change the part of the DSRG model using Class Activation Map(CAM)algorithm combined with VGG16 classification network to using Gradient-weighted Class Activation Mapping(GRAD-CAM)algorithm combined with ResNet101V2 classification network.This paper is called the GRAD-CAM_DSRG model.By changing the VGG16 classification network to the better performance of the ResNet101V2 classification network,more salient features can be obtained.By improving the CAM algorithm to the GRAD-CAM algorithm,the convolutive neural network can be better interpreted.The combination of the two algorithms can expand the regional scope of the initial seed.The problem of seed sparsity faced by DSRG model is solved to some extent.In order to solve the problem of over-growth or under-growth in the process of seed expansion faced by the studied DSRG model,the segmentation network of DSRG model was improved.The DeepLabV2 model in the DSRG model was improved to the DeepLabV3+model.DeepLabV3+model can extract more spatial texture information at more scales than DeepLabV2 model,which improves the accuracy of network segmentation.Combined with the improved part of the classification network,it is called GRAD-CAM_DeepLabV3+_DSRG model in this paper.The segmentation performance of GRAD-CAM_DSRG model,GRAD-CAM_DeepLabV3+_DSRG model and DSRG model were compared on the enhanced SBD dataset of PASCAL VOC 2012.Experimental results of segmentation performance comparison in this paper show that the MIOU value of the GRAD-CAM_DSRG model is 59.5%,which is 0.5%higher than that of the DSRG model,and the MIOU value of the GRAD-CAM_DeepLabV3+_DSRG model is 60.2%.Compared with the MIOU value of DSRG model,the MIOU value is increased by 1.2%,which proves the feasibility of the improved method.