PFL-YOLO: A Lightweight Model for Steel Surface Defect Detection via Partial Frequency-Domain Fusion and Scale-Aware Prediction

Steel surface defects vary widely in scale and often appear as subtle texture changes, making accurate, efficient detection challenging. However, lightweight detectors must preserve these cues while reducing multi-level prediction redundancy and maintaining reliable localization across diverse defect shapes. Therefore, this study proposes PFL-YOLO, a lightweight steel surface defect detector built on YOLO11n. Specifically, the Partial Frequency Adaptive Module (PFAM) applies Frequency Dynamic Convolution (FDConv) to a subset of feature channels while retaining an identity bypass for the remaining channels. This partial-channel design introduces frequency-aware weight adaptation without applying FDConv across the entire feature tensor. In addition, the Lightweight Direct-Residual Scale-Aware Detection Head (LDRSA-Head) shares feature transformations and task predictors across prediction levels, thereby reducing cross-level parameter duplication. Meanwhile, it retains level-specific residual coefficients and output-scaling factors for each prediction level. For bounding-box regression, PFL-YOLO adopts Shape-IoU to penalize center displacement and normalized width and height discrepancies. On the NEU-DET dataset, PFL-YOLO achieved mAP50 of 79.4% and mAP50-95 of 44.7%, exceeding YOLO11n by 1.7% and 1.2%, respectively. On the GC10-DET dataset, the corresponding improvements were 2.1% and 1.1%. Moreover, the model reduced the parameter count by 27.5% and GFLOPs by 19.0%, while achieving 120 FPS on the NEU-DET dataset and 109 FPS on the GC10-DET dataset. In conclusion, these results demonstrate a more favorable accuracy–efficiency trade-off than YOLO11n across both datasets.