PDGA-YOLO: An Improved Steel Surface Defect Detection Network Based on YOLOv8n

Abstract

Steel surface defect detection is essential for ensuring product quality. Industrial defect detection is challenged by redundant spatial computation, loss of deep feature information, and limited local receptive fields. To address these limitations, we propose PDGA-YOLO, a steel surface defect detection network based on YOLOv8n. First, we develop a Partial Convolution and Energy Recalibration C2f (PCER-C2f) module that uses partial convolution (PConv) to reduce redundant spatial computation. It also applies the simple, parameter-free attention module (SimAM) to recalibrate fused features without introducing additional learnable parameters. Second, we design a Dual-Pooling Spatial Pyramid Pooling-Fast (DP-SPPF) module that combines max- and average-pooling information to enrich multiscale contextual representations. Third, a Global Position-Sensitive Attention C2f (GPSA-C2f) module uses position-sensitive self-attention to improve the modeling of long-range dependencies. On the NEU-DET dataset, PDGA-YOLO achieves an mAP@50 of 81.3%, a precision of 76.8%, and a recall of 76.1%. These values exceed those of the YOLOv8n baseline by 1.4, 1.3, and 2.7 percentage points, respectively, while slightly reducing computational cost and parameter count. The results demonstrate the potential of PDGA-YOLO as a high precision solution for steel surface defect detection.