MGA-YOLO: A Surface-Guided Multi-Task Network for Industrial Defect Detection

Abstract Industrial surface defect detection is critical for manufacturing quality control. However, complex mechanical components comprising both rough casting surfaces and precision-machined surfaces pose significant challenges due to the visual similarity between casting textures and actual defects, leading to high false-alarm rates. This paper proposes MGA-YOLO, a lightweight, real-time multi-task network optimized for high-precision defect detection on machined surfaces. Built upon a YOLOv8m shared backbone to ensure computational efficiency, MGA-YOLO introduces three key innovations:(1)A Channel-Alignment Segmentation Decoder integrated with a Boundary F1 Loss to precisely delineate machined regions;(2)A Mask-Guided Residual Attention (MGRA) module that utilizes segmentation features as spatial and channel priors to dynamically enhance feature responses in regions of interest while suppressing casting background interference;(3)A Task-Specific Bias Initialization (TSBI) strategy designed to mitigate task dominance and gradient interference during the early stages of multi-task learning. Experimental results on an industrial dataset demonstrate that MGA-YOLO achieves a 76.7\% mAP@50 and 88.1\% mIoU. Compared to single-task baselines, the proposed model significantly reduces false detections in non-machined areas while maintaining a real-time inference speed of 135 FPS, satisfying the stringent requirements for edge-side deployment in industrial environments.