Establishing the relationship between the feature information of the target objects in the support image and the query image is the mainstream approach of existing few-shot semantic segmentation methods.However,most of these methods do not mine background information to assist the prediction of foreground target information.Therefore,a few-shot segmentation network that interacts with foreground and background features is proposed.Specifically,a new branch is added to the traditional two-branch few-shot segmentation network based on target information to explicitly predict the background region of the query image,and then the predicted background region is used to correct the predicted foreground target region.Meanwhile,in the K-shot scenario,the support bias is alleviated by adaptively adjusting the contribution proportion of a single support image in the support prototype.Finally,experiments are conducted on standard few-shot segmentation datasets.The results on the 1-shot setting of the PASCAL-5i dataset showed that the proposed method achieved 0.4%and 2.8%improvement in mIoU compared to NTRENet and MMNet,respectively.
few-shot learningsemantic segmentationno target region prediction
国家科技期刊平台
NSTL
万方数据
