Background-Weaken Generalization Network for Few-Shot Industrial Metal Defect Segmentation

Abstract

Identifying surface defects in industrial metal fabrication is a vital quality control task that is essential for maintaining product quality and safety. Traditional surface defect identification algorithms depend on a substantial quantity of labeled data, which typically necessitates considerable time and human resources. Additionally, these methods often require retraining when dealing with new types or specific metal defects. This research suggests a new background-weaken generalization network (BGNet) to address these challenges by diminishing the impact of background and improving generalization in the few-shot segmentation of industrial metal defects. BGNet introduces the compressed strengthening (CS) module, center memory (CM) feature fusion module, and cross embedding (CE) module. The CS module consists of a compressed sensing block and a multilevel feature strengthening block, which reduces background interference and enhances the defect foreground through dimensionality reduction and feature enhancement with different receptive fields. The CM feature fusion module activates clear object features by utilizing fuzzy memory features, strengthening the mapping relationship between sets. The CE module mines the relationships between images in different sets through cross-guidance operations, enabling the model to better generalize to new defect classes. The results of the experiments on different datasets show that BGNet delivers the best performance currently available.