Research On Weakly-Supervised Semantic Segmentation Based On Salient Information Fusion

Semantic segmentation is one of the common tasks in the field of computer vision.Research on semantic segmentation under supervised learning is often based on pixel-level labels,which requires a lot of manpower and material resources to manually label.This paper mainly studies the problem of weakly-supervised semantic segmentation based on Image-level tags.In the research of weakly-supervised semantic segmentation,the weakly-supervised localization of target objects in images is one of the main research directions and breakthroughs.Most of the existing segmentation models based on image-level labels use CAM(Class Activation Mapping)algorithm for weakly-supervised localization,and get the seed region of the target.However,the initial seed area obtained by CAM is often very sparse and the coverage area is small.For multi-tasks,the localization effect of multi-target images is not particularly ideal,which will affect the accuracy of the whole segmentation model.The salient information of the image also contains a large number of location information and boundary information of the target object.Experiments show that significant information can well complement weakly-supervised localization information.However,in the current semantic segmentation task,salient information is rarely used for weakly-supervised localization of target objects or only for background localization.Therefore,in this paper,we consider giving full play to the role of significant information and integrating it into the weakly-supervised localization algorithm.On this basis,the precision of weakly-supervised semantic segmentation is further improved.In view of the above problems,the main work of this paper is as follows:(1)Propose our own weakly-supervised localization algorithm SAL-CAM(Salient Class Activation Mapping).In order to solve the problems of scarce localization information and inaccurate localization provided by CAM(Class Activation Mapping)algorithm in semantic segmentation tasks,a Sal-CAM algorithm is proposed.On the one hand,the salient information of the image is fused into the weakly-supervised localization algorithm.On the other hand,the original CAM uses GAP(Global Average Pooling)and adopts the idea of averaging gradients to make sure that the original network structure is not changed.Experiments show that Sal-CAM has higher weakly-supervised localization accuracy and can provide more reliable localization information for semantic segmentation tasks.(2)Pan-CAM(Panda Class Activation Mapping),a weakly-supervised localization algorithm for animal visual interest targets is proposed.Through the first series of panda eye movement experiments in China,we have obtained the traditional features of images that are of visual interest to pandas.Using these traditional features,we have optimized the weakly-supervised localization algorithm(Sal-CAM)and proposed the Pan-CAM algorithm.Experiments show that Pan-CAM algorithm can locate the panda's interested target in the image more accurately(3)On the basis of Sal-CAM,the initial seed required by the semantic segmentation task is obtained.Combined with the idea of seed region expanding SRG(Seeded Region Growing),and at the same time,the loss function considering both foreground and background is adopted to propose its own weakly-supervised semantic segmentation model S-SRG(Salient Seed Region Growing).Experiments show that the S-SRG model has higher precision than the semantic segmentation model of the same type..