To address the issues of inaccurate positioning, weak feature extraction capability, and poor cross-domain adaptability in the detection of surface defects of steel materials, this paper proposes an improved YOLOv11-EMD algorithm and integrates a multi-stage transfer learning framework to achieve high-precision, robust, and low-cost industrial defect detection. Specifically, the InnerEIoU loss function is introduced to improve the accuracy of bounding box regression, the multi-scale dilated attention (MSDA) module is integrated to enhance the multi-scale feature fusion capability, and the Cross-Stage Partial Network with 3 Convolutions and Kernel size 2 Dynamic Convolution (C3k2_DynamicConv) module is embedded to improve the expression of and adaptability to complex defects. To address the problem of performance degradation when the model migrates between different data domains, a multi-stage transfer learning framework is constructed, combining source domain pre-training and target domain fine-tuning strategies to improve the model’s generalization ability in scenarios with changing data distributions. On the comprehensive dataset constructed of NEU-DET and Severstal steel defect images, YOLOv11-EMD achieved a precision of 0.942, a recall of 0.868, and an mAP@50 of 0.949, which are 3.5%, 0.8%, and 1.6% higher than the original model, respectively. On the cross-scenario mixed dataset composed of NEU-DET and GC10-DET data, the mAP@50 was 0.799, outperforming mainstream detection algorithms. The multi-stage transfer strategy can shorten the training time by 3.2% and increase the mAP by 8.8% while maintaining accuracy. The proposed method improves the defect detection accuracy, has good generalization and engineering application potential, and is suitable for automated quality inspection tasks in diverse industrial scenarios.

At present, the detection of steel surface defects is still challenging, because there are some problems in steel products, such as complex background and noise interference, making it difficult to accurately detect complex small targets and great changes in defects at different scales, which directly affects product quality and endangers life safety. To solve the above problems, this paper proposes a steel surface defect detection network based on global attention perception and cross-layer interactive fusion, named GCF-Net. Firstly, this paper proposes an Interactive Feature Extraction Network (IFE-Net), which uses a local modeling feature extraction module to enhance the extraction of local detail features and uses a global attention perception module to capture the global contextual information in the image, thus improving the detection of complex background and noise defects. Secondly, this paper proposes a Cross-Layer Interactive Fusion Network (CIF-Net), which makes up for the fine-grained information lost during the gradual refinement of features through the fusion of adjacent layers, fully integrates shallow and deep features, and at the same time enhances the interaction between different scales by cross-layer fusion, thus improving the recognition ability of defect targets of different scales. Thirdly, the Interactive Fusion Module (IFM) is proposed, which can adjust the importance of each mosaic feature by attention to make efficient use of all feature information and improve the detection of complex background defects. Finally, in order to solve the problems of difficult positioning and inaccurate detection of small targets, this paper aims to strengthen the sensitive loss Q_IOU of small targets and improve the perception of complex small targets in steel defects. Compared with the baseline model, mAP@.5 is improved by 7.0%, 4.4%, and 2.5% on the NEU-DET, PCB, and Steel datasets, respectively, and it is better than all of the comparison models.