Hybrid Feature Enhancement and Adaptive Loss Frame for Few-Shot Semantic Segmentation

Although few-shot semantic segmentation methods have widely studied in computer vision, it still has room for improvement. In this work, we propose to enrich the feature representation with texture information and assign adaptive weights to losses. Specially, we incorporate the texture information obtained by texture enhance module with layer’s features on ResNet, and then get a series hybrid features. The incorporation of texture information enhances the similarity calculation to make the support set guidance more effective. Besides, the proposed adaptive loss makes the network optimize in a better direction. The experiments testify that the proposed method achieves the SOTA on few-shot segmentation dataset such as PASCAL5i and FSS-1000.