Research On Weakly Supervised Semantic Segmentation Method Based On Multi-level Information Fusio

The semantic segmentation task is to assign a semantic category label to each pixel of a given image.Currently,this task has been widely used in various fields such as intelligent security,aerospace,human-computer interaction,it has important social and economic value.In recent years,with the development of deep convolutional neural networks,semantic segmentation networks have achieved significant improvements in accuracy under the supervision of a large number of pixel level labels,but the process of obtaining pixel level labels is time-consuming and laborious.In order to reduce the cost of label acquisition,the weakly supervised semantic segmentation method based on category label has received widespread attention from researchers because it only requires the category label information of the image to segment the image.Due to the pooling operation in classification networks,weakly supervised semantic segmentation methods currently obtain class activation graphs with sparse discriminant regions and fuzzy target edges.For this reason,this article explores the weakly supervised semantic segmentation method based on category label from the perspective of multi-level information fusion,with the specific content as follows:(1)In order to solve the problem of sparse discriminant regions in class activation maps,work 1 of this paper proposes a multi-level online iterative update method for learning class activated maps.The algorithm consists of two parts: a Refinement Gradient Class Activation Maps generation module and a Multilayer Online Accumulation of Class Activation Maps mechanism.The Refinement Gradient Class Activation Maps generation module uses gradient information to generate more accurate class activation maps,while being able to mine more target regions.The Multilayer Online Accumulation of Class Activation Maps mechanism iteratively fuses the class activation maps generated from the characteristic maps of each layer of the backbone network to obtain corresponding online class activation maps,and then combines the online class activation maps of each layer.This accumulation mechanism can effectively supplement some details of the class activation maps.Compared with weakly supervised semantic segmentation algorithm in recent years on the public dataset PASCAL VOC 2012,the results show that this method has significant performance improvements,verifying the effectiveness of this method.(2)In the second work of this paper,a weakly supervised semantic segmentation algorithm based on multi-scale feature fusion is proposed to solve the problem of fuzzy edges of targets.This algorithm can simultaneously capture edge information and salient regions of targets.Firstly,a Multi-scale Class Activation Mapping Learning Mechanism is designed that can generate and fuse multiscale information of images.In this process,Multiscale Attention learning algorithms are used to characterize the correlation between class activation maps at different scales and extract edge information of targets.In addition,this work also designed a kind of class activation mapping fusion algorithm that adaptively combines the results of Multiscale Attention learning algorithms.In the training phase,multiple loss functions are used to perform multiple constraints on the model to ensure the quality of the generated class activation maps.Finally,a large number of experiments have proven the effectiveness of this algorithm.