SGNet: A Multi-Scale Method for Steel Surface Defect Detection

The large variation in target scales within steel surface defect images often results in low detection accuracy. To address this issue, we have developed a multi-scale defect detection method named SGNet. Initially, we introduced a Self-Guided Bi-Level Routing Transformer (SG-BiFormer) as the backbone network of our model. SG-BiFormer utilizes a self-guided strategy to model global information and dynamically reallocates tokens based on the importance of image regions, increasing token quantity in significant areas to preserve fine-grained details and reducing it in less critical areas. This strategy significantly enhances the accuracy and efficiency of the model in handling multi-scale targets. Subsequently, we incorporated Grouped Separable Convolutions (GSConv) to manage local information of multi-scale features. GSConv automatically adjusts the size and shape of the convolution kernels according to different input features. This design not only improves global information processing capabilities but also ensures sensitive and precise detail capture. Lastly, we employed MPDIoU as the loss function for our model, which optimizes corner positioning to tackle the challenges of regression tasks, substantially improving the accuracy of the model in handling complex defects. Experimental evidence confirms the effectiveness of this method. Our SGNet achieved a 77.1% mAP on the NEU-DET dataset, demonstrating outstanding performance.