Few-Shot Segmentation via Information Interaction Enhancement and Multiscale Feature Aggregation

Few-shot segmentation aims to segment unseen classes with few labeled support sets. Most existing prototype-based approaches struggle to handle image pairings with large intra-class differences, and the bias problem leads to capturing seen classes while failing to recognize unseen ones. To mitigate the drawbacks, we propose an information interaction enhancement and multi-scale feature aggregation network (IMNet) for few-shot segmentation. The information interaction enhancement module is designed to capture and leverage the pixel-wise discriminative information, further exploring the correlation between support-query feature pairs. The multi-scale hierarchical decoder employs a hierarchical U-shaped architecture to aggregate and refine features across different scales, enabling effective integration of both coarse-grained and fine-grained representations through its bottom-up pathway. Furthermore, we incorporate a lightweight attention guided module to mitigate bias problem and enhance segmentation precision. Experiments on Pascal-5i and COCO-20i show that IMNet significantly outperforms state-of-the-art methods.