Research On Semantic Segmentation Method Based On Weakly Supervised Learnin

In recent years,with the upgrading of hardware such as GPU,semantic segmentation methods based on deep learning theory have developed rapidly.Semantic segmentation methods are mostly based on fully supervised labels,which are expensive to produce.Therefore,combining semantic segmentation methods with low-level labels has become a new research direction.In terms of the selection of low-level labels,most recent methods adopt class activation mapping based on image-level labels.The related optimization methods designed in this paper aim to mine low-level labels and the supervision information of the network itself.The main work of this paper can be summarized as the following two stages:The first stage is the pseudo annotation generation stage based on equivariant constraints.In this paper,a gated cross-feature attention equivariant constraint method is proposed.In order to alleviate the inaccuracy and incompleteness of the CAM method,three optimization methods are proposed in this stage,including the multi-twin equivariant constraints,the gated RGB feature fusion module,and the cross-feature attention module.In a series of experiments,pseudo-annotations generated by this method achieved m Io U values of 64.9% and 66.0% in VOC 2012 val set and test set,and outperform other state-of-the-art algorithms.The second stage is the semantic segmentation stage based on pseudo-annotations.This paper proposes a modified Deep Lab-ASPP method based on pseudo-annotations.How to effectively train the segmented network by using inaccurate pseudo-annotations also belongs to weakly supervised learning.In this stage,the multi-scale boundary feature mining module and the modified self-regulation constraint module are designed to improve the accuracy of semantic segmentation network under pseudo-annotation supervision.In VOC 2012 val set and test set,this method outperforms other existing state-of-the-art algorithms with m Io U values of 65.9% and 66.7%,and outperforms pseudo-annotations by 1.0% and 0.7%.In COCO data set,this method achieves the m Io U value of 33.0%,slightly ahead of CONTA model,and outperforms pseudo-annotations by 0.6%.Compare with pseudo-annotations,the boundary generated by this method is closer to the boundary of ground truth.