Improved YOLOv11 for Steel Surface Defect Detection Based on PMDT-YOLO

To address the issue of low detection accuracy caused by the large number of tiny defects, significant scale variations and complex backgrounds in steel surface defect detection, we propose a detection algorithm Pinwheel-shaped Convolution-based YOLOv11 using Mixed Aggregation Network and Dynamic Tanh (PMDT-YOLO). First, the Pinwheel-shaped Convolution (PConv) is introduced, which utilizes Gaussian distribution features to improve the model's capacity to detect tiny defects. Next, the Mixed Aggregation Network (MANet) is incorporated in the neck of the model to integrate features from different levels, enhancing the semantic depth of feature extraction. Finally, we introduced DyT (Dynamic Tanh), incorporating learnable DyT into the C2PSA module to enhance the model's adaptability to varying environments through nonlinear transformations. Experimental findings indicate that on the NEUDET public dataset, the improved model attains an mAP of 81.7%, marking a 2.3% enhancement over the original YOLOv11 model, with Precision rising by 4.4%. Similarly, experiments on the GC10-DET dataset demonstrate that the PMDT-YOLO shows considerable robustness, indicating it can effectively detect surface defects in steel.