Image Semantic Segmentation Method Based On Weakly Supervised Learning

The image semantic segmentation is one of the fundamental research tasks in the field of computer vision to achieve pixel-by-pixel classification of input images.The semantic segmentation task under fully supervised conditions achieves better results by virtue of pixel-level annotation with real labels,but obtaining fine pixel-level annotation is time-consuming and costly.Therefore,research on semantic segmentation under weakly supervised conditions has emerged to overcome the problem of fully supervised reliance on pixel-by-pixel labeling by using weak labels such as points,scribbles,bounding boxes,and image levels that are relatively easy to obtain in weakly supervised semantic segmentation task.In view of the fact that weakly supervised semantic segmentation can guarantee the accuracy of semantic segmentation while reducing the workload of data annotation,this thesis is devoted to semantic segmentation under weakly supervised conditions,and the research work is carried out using the most easily accessible image-level labels among weak labels,and two methods are proposed to further improve the performance of weakly supervised semantic segmentation,and the main work is summarized as follows:Firstly,to address the problem that the class activation map only activates part of the discriminative regions that cannot meet the target localization demand,a weakly supervised semantic segmentation method with improved attention map and random walk and adaptive refinement of diffusion pseudo label by image slicing and region restraint is proposed.The method firstly drives the classification network to focus on the discriminative regions in different parts of the image through a pre-processing step of image slicing;secondly,the region restraint module is placed in the classification network to constrain the discriminative regions of the target object during feature extraction to activate the surrounding non-discriminative regions and generate dense class activation maps;finally,affinity random walk strategy is used to further refine the semantic propagation among the same classes.Finally,the affinity random walk strategy is used to achieve semantic propagation among the same classes,and the adaptive refinement module with image color information and spatial location information is used to refine the pseudo label and use it as supervised information to train the segmentation network.The proposed method was experimentally validated on the VOC 2012 dataset,and 66.4%segmentation results were obtained.The method effectively improved the quality of the class activation map and achieved better segmentation performance.Secondly,to address the problem of the gap between image-level labels and pixellevel segmentation results,a weakly supervised semantic segmentation method based on attention mechanism and boundary exploration is proposed in order to learn pixel-level semantic knowledge from image-level class labels.The method first obtains dense localization maps by using a classification network with a region restraint module,and then obtains the remote context information of the image features using an enhanced selfattention module;secondly,it embeds the structural information of the target into the segmentation model using a boundary exploration module,refines the pseudo label using boundary-guided smoothing loss,and makes the pseudo label boundary-aware;finally,to further improve the quality of the supervised information provided to the segmentation network,the pseudo label indicates reliable foreground information,the saliency map indicates background information,and improves the saliency fusion to obtain pixel-level pseudo label,and the segmentation result is obtained through the segmentation network.On the VOC 2012 dataset,the method achieved 67.2% of mIoU values,which is better than other weakly supervised semantic segmentation methods.