DG-Net: a steel surface defect detection method using dynamic receptive field and gradient attention

Abstract The surface defects of steel materials are diverse in type and shape. Existing detection methods, while improving overall accuracy, often lead to a decline in the detection precision for certain defect categories. To balance the detection performance across multiple defect types, this study proposes the DG-Net detection model. First, the model incorporates the DynamicSPPF module, which dynamically adjusts the pooling kernel to accommodate defects of varying scales, significantly enhancing the ability to extract multi-scale features. Second, to address the limitations of the Bbox Loss Path in traditional head structures for fine defect localization, a Gradient Enhancement Module (GEM) is introduced and integrated into the Bbox Loss Path. This module improves the sensitivity to fine defects by weighting the gradient information of the input feature map. Finally, a Gradient-Driven Attention-Enhanced Detection Head (GDAEDH) replaces the traditional detection head. GDAEDH further strengthens the expression capability of local features by introducing the ACmix attention mechanism after the GEM module, thereby effectively improving defect localization and recognition accuracy. Through comparisons with over thirty mainstream algorithms, DG-Net achieved a peak detection accuracy of 82.7% on the NEU-DET dataset, demonstrating outstanding generalization capabilities on the GC10-DET and APSPC datasets.