FPIseg: Iterative segmentation network based on feature pyramid for few‐shot segmentation

Abstract Few‐shot segmentation (FSS) enables rapid adaptation to the task of unseen‐classes object based on a few labelled support samples. Currently, focal point research in FSS field is align features between and query images, aiming improve performance. However, most existing methods implement such support/query alignment by solely leveraging middle‐level feature for generalization, ignoring category semantic information contained high‐level feature, while pooling operation inevitably lose spatial feature. To alleviate these issues, authors propose Iterative Segmentation Network Based Feature Pyramid (FPIseg), which mainly consists three modules: Fusion Module (FPFM), Region Enhancement (RFEM), Optimization (IOSM). Firstly, FPFM fully utilizes foreground from image under multi‐scale, multi‐level backgrounds. Secondly, RFEM enhances detail aligned generalization ability. Finally, ISOM iteratively segments optimize prediction result Extensive experiments PASCAL‐5 i COCO‐20 datasets show that FPIseg achieves considerable performance both 1‐shot 5‐shot settings.