Few-Shot Segmentation Network Robust to Background Interference

Publication Name

Proceedings - 2023 IEEE International Conference on High Performance Computing and Communications, Data Science and Systems, Smart City and Dependability in Sensor, Cloud and Big Data Systems and Application, HPCC/DSS/SmartCity/DependSys 2023

Abstract

Few-shot segmentation has gained significant attention owning to the effectiveness in segmenting unseen classes with a few annotated images. However, there exist two challenges in previous works. 1) They focus on extracting foreground features of support images to guide the segmentation of unseen classes, which causes the loss of useful information and obtains a limited representation of the overall context. 2) They inevitably produce a bias towards base (seen) classes due to the meta-training on the base dataset. That is, the segmentation performance of models can not be guaranteed when predicting images whose backgrounds are similar to classes in the base dataset. To this end, a few-shot segmentation network robust to the background interference (RB-net) is proposed. Specifically, RB-net utilizes middle layers of feature extractors to extract multi-level representations for enhancing the generalization of features. Then, a self-guided prototype correlation learning is designed via modelling correlation matrix encoding between support-query image pairs, to learn foreground and background prototypes with the help of the pyramid pooling encoder, which effectively guides the segmentation of the query image. Meanwhile, a base-class learner based on meta learner is added to predict backgrounds similar to base classes in the query image, for alleviating the bias problem of seen classes. They seamlessly cooperate to suppress the interference of background noises and enhance the segmentation to unseen classes with blurry edges and complex backgrounds. Finally, extensive experiments demonstrate the superiority and effectiveness of RR-net.

Open Access Status

This publication is not available as open access

First Page

131

Last Page

137

Funding Number

62076047,62006035

Funding Sponsor

National Natural Science Foundation of China

Link to Full Text

Share

COinS