液晶与显示2024，Vol.39Issue(10)：1421-1430,10.DOI:10.37188/CJLCD.2024-0074
基于双重聚合和自合并网络的小样本图像语义分割
Bi-aggregation and self-merging network for few-shot image semantic segmentation
摘要
Abstract
Few-shot image semantic segmentation is a very challenging task that attempts to segment objects of new classes using only a few labeled samples.The mainstream methods often have problems of low discriminative feature and prototype deviation.To alleviate these problems,a new few-shot image semantic segmentation method based on a bi-aggregation and self-merging network is proposed,which can fully mine the similarity of features and reduce prototype bias.Firstly,we propose a feature-mask bi-aggregation module to provide global semantic information for the feature aggregation and mask aggregation by constructing a dense similarity relation between the support features and the query features covering all spatial locations.Specifically,an enhanced feature and an initial mask with guiding information can be obtained for the query image by performing feature and mask bi-aggregation on the similarity matrices.Then,a self-merging decoder is proposed,which reduces the prototype bias by adding the initial mask-based self-prototype with the known support prototypes,and conveys rich category semantic information to the decoder by fusing the merged prototype with the enhancement feature.Finally,the prediction results obtained by the decoder are further optimized by the prediction results of the base classes.The mIoU values of our method on the dataset PASCAL-5i achieve 68.3%and 71.5%in the 1-shot and 5-shot cases,respectively,and on the dataset COCO-20i achieve 46.5%and 51.4%in the 1-shot and 5-shot cases,respectively,which is superior to the segmentation performance of the mainstream methods,and can segment the target region of the new class more accurately.关键词
小样本图像语义分割/特征相似性/双重聚合/类内差异性/自合并Key words
few-shot semantic segmentation/similarity of features/bi-aggregation/intra-class diversity/self-merging分类
信息技术与安全科学引用本文复制引用
刘玉,于明,朱叶..基于双重聚合和自合并网络的小样本图像语义分割[J].液晶与显示,2024,39(10):1421-1430,10.基金项目
国家自然科学基金青年项目(No.62102129) （No.62102129）
河北省自然科学基金(No.F2021202030)Supported by Youth Program of National Natural Science Foundation of China(No.62102129) （No.F2021202030）
Natural Science Foundation of Hebei Province(No.F2021202030) （No.F2021202030）
