Frequency separation-based few-shot segmentation

In recent years, Few-shot Semantic Segmentation (FSS) has aimed to segment unseen class targets using a small number of labeled samples while leveraging high-level semantic features to address spatial inconsistencies between query and support targets. Traditional approaches often rely on prototype vectors and metric functions for feature interaction but fail to comprehensively capture all the features of the support images. To enhance detail extraction and feature alignment, we propose the High-Low Frequency Feature Fusion Attention Module (HLSFA). This module separates high-frequency and low-frequency components of the features. It then computes attention weights for the high-frequency components independently, significantly improving the representation of target region features. Additionally, we introduce the Recursive Cosine Fusion Module (RCFM), which enhances the representation of support and query features using cosine similarity and recursive enhancement mechanisms. Finally, the SASPP module is employed to fuse multi-scale features, further improving segmentation accuracy. Our approach achieves notable progress on the PASCAL-5i\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5<^>i$$\end{document} and COCO-20i\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$20<^>i$$\end{document} benchmark datasets, reaching mIoU scores of 69.4, 70.2, 55.9, and 57.6%, respectively, showing significant improvement over existing methods. The code is publicly available at https://github.com/zxmyyds/FSFNet.