Surface defect detection YOLO based on prototype enhancement

Abstract

With the rapid development of deep learning, industrial defect detection has achieved significant improvements in performance and accuracy. However, most existing methods depend on large-scale annotated datasets and substantial computational resources, which limits their applicability in industrial scenarios where defect samples are scarce and expensive to label. To address this issue, this paper proposes a prototype-enhanced YOLO-based defect detection model for few-shot industrial applications. Prototype prior information is introduced to guide feature learning and improve detection performance under limited training samples. Specifically, a prototype learning module extracts class prototypes from the global structural information of normal samples to enhance the discriminative capability of defect features. A prototype enhancement module further generates defect prototypes to guide the model toward critical defect regions while providing auxiliary category priors for the detection branch. In addition, a consistency constraint loss is designed to incorporate prototype branch information into model optimization, thereby improving feature consistency and generalization capability. Experimental results demonstrate that the proposed method outperforms mainstream defect detection models, validating its effectiveness in few-shot industrial defect detection tasks.