Weakly Supervised Semantic Segmentation Based On Federated Learning

Image semantic division technology is one of the core tasks in the field of computer vision.Semantic segmentation technology is widely used in various fields of production and life,such as medical analysis,urban intelligent transportation,and scene understanding.The achievement of the semantic segmentation is greatly achieved by the training network using accurate to pixel-level labels,but the acquisition of pixel-level labels is more difficult,making the cost of network training expensive.The recently proposed weakly supervised semantic segmentation method is to reduce the accuracy of the label and use image-level labels to train the network to reduce the training cost.By continuously digging the connotation information of image-level labels,the semantic segmentation model can be ensured while reducing the cost of labeling data.Based on the problem of weakly supervised semantic segmentation,this thesis uses the goal of proposing a general scene segmentation model,ensuring model segmentation accuracy,and improving the target positioning ability under weakly supervised.In depth research on the semantic segmentation of weakly supervised,and designing universal weakly supervised semantic segmentation networks.The main content of this thesis is summarized as follows:(1)In response to the condition of weakly supervised,the activation diagram only focuses on the problem of the most distinguished area in the image,and proposes a class diffusion activation graph method integrated into the attention matrix.The new feature extraction network is added to the feature extraction process,and the global pixel attention matrix is obtained.The matrix has the attention relationship between the pixel pair of pixels.The flexible attention diffusion module proposes to the spread of pixels with high similarity,which can improve and expand and expand the activation area of the original class activation diagram.By accumulating the class activation diagram obtained from different training stages,the activation area will cover as much as possible.The cumulative process based on dynamic weights can effectively reduce the activation area of the error generated by the network.Experimental verification on the method mentioned on VOC 2012 and MS COCO data sets,obtained 67.4% and36.8% of the segmentation results,which improved the target positioning and segmentation capabilities of the network under weakly supervised.method.(2)In response to the problems of weakly supervised semantic segments,incomplete division of target areas and serious target boundaries,etc.,and proposes to be a weakly supervised semantic segmentation model based on multi-tasking joint learning.This model has a total of three learning tasks.The first task is to learn and classify the activation diagram for input images for input images.The significant detection of comparison learning improves the ability of the network to perceive the network;the third task is the final image segmentation task.Significant detection and segmentation tasks use similar learning modules to perform semantic information communication,and optimize the significant diagrams and segmentation diagrams.Obtained 68.1% and 36.9% segmentation on the VOC 2012 and MS COCO datasets,respectively,and the segmentation accuracy was better than other models of the same period.