Multi-information fusion based frequency-aligned network for few-shot segmentation

Few-shot segmentation endeavors to extract knowledge from limited samples and then use it to segment images in novel categories. However, owing to the paucity of source data, extracted knowledge is often insufficient to handle variations in local features or issues such as perspective distortion. To address this problem, we propose MFANet, which effectively leverages textual and instance information to achieve accurate segmentation. Specifically, the network includes a frequency interaction module that integrates features across different frequency bands, alleviating discrepancies between frequency-specific features and helping identify defective ones. Subsequently, an information fusion module combines textual information with feature representations to mitigate challenges such as ambiguous local features or perspective distortions. Moreover, to ensure consistent learning between text and image modalities, we introduce a multimodal alignment module that maps both modalities into a unified embedding space. Extensive experiments on P A S C A L − 5 i and C O C O − 2 0 i demonstrate that MFANet achieves superior performance.