面向工业表面缺陷的低复杂度检测网络OA

To address industrial scenarios characterized by"small but numerous defects,complex backgrounds,and limited computing power",this paper proposes the YOLO-DAFE architecture.The framework first designs FD-DConv for frequency-domain decoupling and reweighting to capture subtle defects,then introduces C2f-OREPA by training multi-branches and reparameterizing inference to overcome low FLOPs constraints,and finally combines Focaler and IoU for joint optimization of localization and classification.On the enhanced NEU-DET dataset,YOLO-DAFE achieves 94.3%mAP with only 6.2G FLOPs,reducing computational cost by 1.9G while improving accuracy by 1.8%compared to the baseline.On the generalization dataset GC10-DET,it achieves 1.3 percentage points higher mAP and 1.2 percentage points higher mAP@0.5:0.95 compared to the baseline.Final validation through Jetson Orin Nano Super resource-constrained devices demonstrates that both datasets match PC-level accuracy and maintain the highest mAP50 across all models,showcasing strong adaptability for industrial applications.Additionally,it achieves the fastest inference time under GPU and FP16 modes with quantization degradation<1%,balancing high precision,low latency,and easy deployment to meet real-time industrial defect detection requirements.