Adaptive similarity-guided self-merging network for few-shot semantic segmentation

Abstract

Few-shot Semantic Segmentation (FSS) attempts to segment the new category with only a few labeled samples, presenting a significant challenge. Existing approaches primarily focus on leveraging category information from the support set to identify objects of the new category in the query image. However, these models often struggle when confronted with substantial differences between paired images. To address issues stemming from scenario differences and intra-class diversity, this paper proposes an adaptive similarity-guided self-merging network. Firstly, style differences of multi-level features are introduced to alleviate the network's sensitivity to scenario variations and learn an adaptive weight for the K-shot scheme. Secondly, a feature-mask bi-aggregation module is designed to learn an enhanced feature and an initial mask for the query image. Within this module, dynamic correlations cover all the spatial locations, providing global information crucial for feature and mask aggregation. Subsequently, a self-merging module is proposed to alleviate prototype bias. It merges a self-prototype derived from the initial mask with an adaptive weighted support prototype obtained from K support images. Finally, the target object is segmented using the enhanced feature and merging prototype, and segmentation results are further refined by predictions of base categories and an adjustment factor derived from multi-level style differences. The proposed method achieves 69.1% (1-shot) and 72.3% (5-shot) mIoU on the PASCAL-5i dataset, and 47.4% (1-shot) and 52.1% (5-shot) mIoU on the COCO-20i dataset. These results demonstrate state-of-the-art segmentation performance compared to mainstream methods.© 2017 Elsevier Inc. All rights reserved.