DFE-Net: Dual-Stream Gated Cross-Modulation and Spatial-Frequency Detail Compensation for Steel Surface Defect Detection

Surface defects on hot-rolled steel strips are often low in contrast, resemble the background, and co-occur. This paper presents DFE-Net, a refinement of YOLO11n for steel surface defect detection. A defect-aware dual-stream gated cross-modulation module (DA-GCM), derived from the FCM block, lets a semantic stream and an appearance stream modulate each other through a dual-pooling channel attention and a multiscale strip spatial attention and combines them with a pixel-wise gate. A spatial-frequency detail-compensation pyramid (SFDP) injects space-to-depth P2 features into the P3 fusion node, refines them with an FFT-based large-kernel block, and aggregates the inputs of every node with an adopted Global Modulated Fusion (GMF) module. A shared-stem parameter-efficient detection head (SSP-Head) shares a grouped-convolution stem between regression and classification. On NEU-DET, over five paired seeds, DFE-Net reaches 74.27 ± 0.81% mAP@0.5 and 41.83 ± 0.55% mAP@0.5:0.95 with 2.24 M parameters, 0.91 and 0.58 points above YOLO11n; the 95% confidence intervals include zero, although mAP@0.5:0.95 is higher in all five seeds and in all five folds of a cross-validation, a consistent direction that tests corrected for the overlap of the cross-validation training sets do not confirm. DFE-Net is significantly more accurate than YOLOv8n and YOLO26n but on par with YOLOv9t. It has 13.2% fewer parameters than YOLO11n, but its computation rises from 6.3 to 7.4 GFLOPs and its batch-1 latency from 5.63 to 14.31 ms on a desktop GPU and from 10.15 to 28.73 ms on a Jetson Orin Nano. DFE-Net is thus a compact detector with a modest accuracy gain; its robustness and suitability for online inspection require further validation.