Edge-YOLO: Lightweight Multi-Scale Feature Extraction for Industrial Surface Inspection

Abstract

The increasing complexity of industrial quality control necessitates advanced defect detection systems capable of identifying small-scale surface anomalies with high precision. While deep learning methods, particularly You Only Look Once (YOLO) architectures, have shown promise in object detection, they frequently encounter limitations when detecting minute defects, primarily due to inadequate feature retention in shallow network layers and computational constraints in industrial settings. This study introduces Edge-YOLO, an innovative network architecture that enhances defect detection through a novel edge-enhanced backbone (Edge-backbone). The Edge-backbone systematically strengthens edge feature extraction and retention through three synergistic components: an Edge-Sensitive (EdgeS) module for optimized feature initialization, a Cross-Stage-Partial Edge Enhancement (C3E2) module that integrates edge information across network stages, and a Multi-Scale Dilated Convolution (MSDC) module that efficiently fuses multi-scale features through weight-sharing. These modules work in concert to create a comprehensive edge-aware feature extraction pipeline. Comprehensive evaluation across three industrial datasets demonstrates Edge-YOLO’s effectiveness, achieving mean Average Precision (mAP@50) improvements of 2.1%, 3.5%, and 10.3% on NEU-DET, GC10-DET, and GEAR-DET datasets respectively, compared to YOLO11. Notably, Edge-YOLO reduces computational complexity by 69.8% while maintaining competitive accuracy, making it particularly suitable for real-time industrial applications requiring both precision and efficiency.