Abstract: In Semantic segmentation aims to assign pixels in an image to different semantic categories. Traditional methods rely heavily on data to learn features and semantic information, and performance drops when there is insufficient labeled data. This paper studies few-shot semantic segmentation to achieve excellent segmentation and generalization capabilities on a smaller dataset. Existing few-shot semantic segmentation methods commonly use simple fusion strategies to achieve information transmission across branches, which are unable to capture complex feature interaction information. Therefore, this paper proposes a mutually guided conditional coupling network to efficiently fuse the support-query branches. The network designs a coupling information generation module, which focuses on the regions in the support feature highly similar to the query feature, so that each pixel in the query feature can adaptively fuse pixel-level information from the support feature. Additionally, to reduce the negative impact of the query image background on the fusion effect, a conditional coupling module is also designed. In this module, the generated conditions can be used to suppress information absorption in the query image background regions, effectively avoiding interference from background information on the results. This paper conducts extensive experimental verification and analysis on PASCAL-5ⁱ and COCO-20ⁱ, and the experimental results show significant improvements relative to baseline methods.