Abstract: Steel surface defects exhibit diverse morphologies, complex structures, a high proportion of small targets, and susceptibility to interference from environmental factors, while existing defect detection models suffer from complex structures, large parameter counts, and poor detection accuracy and real-time performance. To address these issues, a lightweight and efficient steel defect detection algorithm (MHTD-YOLO11n) based on YOLO11n is proposed in this studyly. Firstly, a multi-scale grouped dilated convolution (MSGDC) module is introduced in this method, in which grouped convolutions with different dilation rates are integrated to achieve multi-scale feature fusion and enhance the detection capability for various types of defects. Subsequently, a Hierarchical Reciprocal Attention Mixer (H-RAMi) module is incorporated to compensate for pixel-level information loss caused by downsampled features. A C2PSA_TPA module is then designed, in which the KV cache size during inference is significantly compressed by leveraging Tensor Product Attention (TPA). Finally, the feature interaction module (C3K2_DFF) is reconfigured to enable the network to effectively combine multi-scale information under a larger receptive field, promoting improvements in both detection accuracy and speed.Experimental results show that compared with the YOLO11n algorithm, the mAP value and recall rate of the MHTD-YOLO11n algorithm are increased by 4.3 and 9.1 percentage points respectively, a detection speed of 258.3 frame/s is achieved, the parameter count and computational volume are reduced by 1.42×10⁶ and 3.4×10⁹ respectively, and the dual requirements of high accuracy and real-time performance in industrial quality inspection scenarios are met.

Key words: computer image processing, steel surface defects, defect detection, object detection, YOLO11n, attention mechanism