PPMA-YOLO: A Method for Metal Surface Defect Detection

Abstract

Metal sheets are vital to core industries like aerospace, automotive, and bridge construction, where their quality directly determines the safety and lifespan of final products. During production, fac-tors like material impurities and equipment wear often cause surface defects such as stains, scratches, and cracks, which can lead to structural failure and corrosion. As manufacturing stand-ards rise, detecting these defects has become essential for maintaining quality and optimizing production processes. To address this, this paper introduces a detection method called PPMA-YOLO. We first replaced the standard YOLOv8 backbone with a lightweight PP-LCNet struc-ture to make the model faster and more efficient, using the H-Swish activation function and adding an SE attention module with large-kernel convolutions to balance precision and speed. In the neck section, we improved the Multi-Contextual Attention (MCALayer) and combined it with multi-scale feature fusion and small residual blocks to create the MAFR module. This module, embedded in the detection head, helps the model capture refined features while avoiding common training issues like gradient explosion. Finally, we updated the SIoU loss function with an adaptive weight and a shape-matching cost, allowing the model to better locate defects of various sizes and shapes. Tests on the NEU-DET dataset show that our method is highly effective; with a model size of only 4.8 MB, it achieved an 83.87% [email protected]. Compared to existing methods, PPMA-YOLO is both more accu-rate and faster, making it an ideal solution for real-time defect detection in industrial settings.