RAGA-YOLO: Enhancing global structural perception for accurate and efficient surface defect detection

Abstract Surface defect detection plays a critical role in quality control of industrial products, particularly in steel manufacturing. However, existing deep learning methods often suffer from high computational costs, limited generalization, and constrained accuracy in complex defect scenarios. To address these challenges, we propose a lightweight detection model, RAGA-YOLO, based on an improved YOLOv11-n. The model incorporates a novel lightweight neck structure, CCFM-PAN-FPN, which combines channel alignment with efficient downsampling to significantly reduce both Params and GFLOPs. Additionally, the Relation-aware Global Attention module is introduced to enhance global information modeling, thereby improving the model’s ability to detect cross-regional defects. Moreover, the C3kDC_FFM module, composed of the DCbottleneck and the Feature Fusion Module (FFM), effectively enhances contextual representation and improves the integration of multi-scale features. Furthermore, a dilated convolution layer is incorporated into the detection head to form a new Dilated head, which enlarges the receptive field and enhances the model’s ability to recognize small-scale defects. On the Northeastern University Steel Surface Defect Detection (NEU-DET) dataset, RAGA-YOLO achieves a 2.1% increase in mAP, substantially reduces multiple Toolkit for Identifying Detection and segmentation Errors (TIDE), and decreases the number of parameters by 15.3% while maintaining unchanged GFLOPs. Experiments on the Metallic Surface Defect Detection with 10 classes (GC10-DET) dataset and Anomaly Surface Visual Defect of washer dataset further demonstrate its excellent generalization capability and robustness to scale variations, highlighting its promising potential for practical industrial applications. In addition, experiments conducted on the PCB dataset further verify the superior performance of the Dilated head in recognizing small-scale defects.