基于DCD-YOLO的钢材表面缺陷检测算法

Steel Surface Defect Detection Algorithm Based on DCD-YOLO

摘要: 本文针对钢材表面缺陷检测中多尺度特征适应性不足、上下文信息融合不充分、变形缺陷捕捉能力有限的问题，提出一种基于YOLOv11n的改进模型DCD-YOLO。首先，设计动态混合卷积模块(DCMB)，替换主干网络C3k2模块的Bottleneck部分，通过动态卷积核权重机制自适应调整卷积核权重大小，增强对多尺度缺陷的特征提取能力；其次，设计了上下文引导的空间特征重构金字塔网络(CGRFPN)，通过矩形自校准模块(RCM)与金字塔上下文提取模块(PCE)，加强模型对缺陷前背景的建模能力，提升缺陷与复杂背景的区分度；最后，通过引入可变形注意力机制(DAttention)替换PSA模块中的固定注意力机制，完成注意力的动态采样，强化了对变形缺陷的适应性。实验结果表明，改进后的模型在GC10-DET数据集上的mAP@0.5达到66.9%，较原YOLOv11n提升3.3%。同时，模型检测精度与召回率分别提升1.7%和2.8%，有效解决了多尺度、背景抑制等检测难题，满足了工业场景对准确性与召回率的要求。

Abstract: This paper addresses the issues of insufficient multi-scale feature adaptability, inadequate contextual information fusion, and limited capability in capturing deformed defects in steel surface defect detection by proposing an improved model, DCD-YOLO, based on YOLOv11n. First, a Dynamic Convolutional Mixed Block (DCMB) is designed to replace the Bottleneck part of the C3k2 module in the backbone network. Through a dynamic convolution kernel weight mechanism, it adaptively adjusts the convolution kernel weights, enhancing feature extraction capabilities for multi-scale defects. Second, a Context-Guided Spatial Feature Reconstruction Pyramid Network (CGRFPN) is designed. By using the Rectangular Self-calibration Module (RCM) and Pyramid Context Extraction Module (PCE), the model’s ability to model defect foreground and background is strengthened, improving the distinction between defects and a complex background. Finally, by introducing a Deformable Attention mechanism (DAttention) to replace the fixed attention mechanism in the PSA module, dynamic sampling of attention is achieved, enhancing adaptability to deformed defects. Experimental results show that the improved model achieves a mAP@0.5 of 66.9% on the GC10-DET dataset, an increase of 3.3% compared to the original YOLOv11n. Meanwhile, the model’s detection precision and recall increased by 1.7% and 2.8%, respectively, effectively addressing detection challenges such as multi-scale defects and background suppression, meeting industrial requirements for accuracy and recall.