<p>To address the issues of high false detection rate, missed detection rate, and low efficiency in small-object defect detection under complex backgrounds, this paper proposes an improved surface defect detection method based on the YOLOv11 algorithm. To enhance the extraction of global contextual information, a module combining CMUNeXt—an efficient medical image segmentation network based on large kernels and skip fusion—with the Efficient Multi-scale Attention (EMA) mechanism is constructed. This module improves the expressiveness of key features during multi-scale feature extraction. Additionally, a Large Separable Kernel Attention (LSKA) module is incorporated at the end of the backbone network, where Ghost convolution replaces standard convolution to further enhance small-object feature extraction while reducing computational cost. In terms of structural improvements, the modified YOLOv11 network adds a high-resolution detection head and removes the large-object detection layer to improve small-object detection precision and optimize computational efficiency. Additionally, to resolve issues of gradient vanishing and insufficient weight optimization in small-object detection, a new bounding box loss function Dpshape-IoU is proposed. Experimental results show that the proposed method improves mAP by 3%, reduces model parameters by 17.7%, and achieves a better balance between detection accuracy and computational efficiency.</p>