Weakly-supervised Image Semantic Segmentation Algorithm Based on Candidate Regions
The existing weakly-supervised semantic segmentation methods rely on initial response and classification task,on-ly focus on distinguishing target object area,and cannot optimize loss function through complete area.This paper presents a seman-tic segmentation algorithm based on candidate regions for weakly-supervised semantic segmentation of image-level annotation data.In this algorithm,mixed data enhancement scheme is first introduced in the classification network,then the corresponding strategy is formulated to cluster image features,subclass targets are constructed and subclass tags are generated,so that the training process is not only dependent on distinguishable object areas.Comprehensive experiments and analyses are carried out on the PASCAL VOC 2012 dataset,the algorithm shows good performance compared to other weakly supervised semantic segmentation algorithms.By us-ing the method of mixed data enhancement and self-supervised candidate regions generation,the original image produces a more complete response map,which improves the Intersection over Union(IOU)by 2.1%and improves the performance of the final seg-menting network.
weakly-supervised learningimage semantic segmentationmixed data enhancementcandidate regions genera-tion
国家科技期刊平台
NSTL
万方数据
