Machine vision-based surface defect testing of industrial materials using a lightweight multiscale attention fusion network

Abstract

Surface defect detection is essential for maintaining quality in industrial products, but in real industrial applications it faces a dilemma among accuracy, processing speed, and model lightweight. To address this challenge, we propose a method named global attention mechanism for deeply separable YOLO (simply denoted as GMD-YOLO) network. It integrates global context and local information in a multi-scale attention fusion module to improve the recognition of small defects, and an enhanced depth separable convolution to reduce computational cost. The model is evaluated in three typical scenarios individually involving pavement, wood surface and steel surface defects. The accuracy, processing speed, and model lightweight are quantitatively analyzed using mean accuracy, frame rate, and floating-point operations, respectively. Results show that GMD-YOLO not only maintains real-time performance but also outperforms the common methods in the YOLO family in all three metrics. High-resolution thermograms reveal the robust decision basis of the model, and ablation experiments confirm the synergistic gain effect between the sub-modules.