GLA-YOLO: A Lightweight YOLO-Based Network for Industrial Surface Defect Detection

Abstract Surface defect detection is essential to quality assurance in modern manufacturing, affecting product reliability, production efficiency, and operational safety across diverse industrial sectors. Despite significant advances in deep learning-based object detection, industrial defect inspection remains uniquely challenging because defects exhibit extremely low contrast against textured backgrounds, highly irregular morphologies, severe scale variations, and marked class imbalance in which critical defect types may appear in fewer than twenty training samples. Deployment on resource-constrained edge devices in factory environments additionally demands real-time inference that many high-accuracy architectures cannot deliver. This paper introduces Gated Local Attention YOLO (GLA-YOLO), a lightweight architecture built upon YOLO11, featuring three innovations: Lightweight Fusion Convolution (LFConv) for efficient shallow feature extraction with channel attention, a Dynamic Local Attention Fusion Block (DLABlock) integrating local depthwise convolution with Performer-inspired linear attention, and DLAC3k2/C2DLA composite modules that embed DLABlock into the YOLO CSP framework. On a self-collected industrial defect dataset (4 categories, 1,367 instances) and the GC10-DET benchmark (10 categories), GLA-YOLO achieves mAP@50 of 90.21% and 80.59%, improving upon YOLO11n by 3.15 and 4.40 percentage points respectively, with only 2.70M parameters and 6.2 GFLOPs, while delivering 168.57 FPS inference and establishing state-of-the-art performance among nano-scale detectors.