FBINet: Few-Shot Semantic Segmentation With Foreground and Background Iteration

Defect detection methods based on few-shot segmentation are becoming more and more popular in industrial applications, and few-shot segmentation methods need to use only a limited number of densely labeled samples to segment objects of unseen classes. However, most existing few-shot semantic semantic segmentation (FSS) approaches primarily focus on either foreground knowledge or background knowledge of an image, both of which are indispensable. Furthermore, relying solely on matching support prototypes may lead to incorrect segmentation regions due to appearance variations between support and query images. To address these issues, we propose a novel framework called the foreground and background iteration network (FBINet) that integrates both foreground and background semantic information. The framework comprises a query learning prototype refinement module, which employs a query feature map as a query prototype to guide target region segmentation. In addition, a query prototype activation module (QPAM) and target-matching prototypes extracted from supporting images are utilized to enhance foreground semantic information. We introduce a background prototype iteration module (BPIM) that iteratively removes the background region of the query image. Through extensive testing on the PASCAL- $5^{i}$ and COCO- $20^{i}$ datasets, we demonstrate the effectiveness of our approach.