Channel spatial semantic attention and grouped dilated fusion for real-time steel surface defect detection

Reliable detection of steel surface defects is essential for ensuring product quality and production safety. However, existing detectors struggle to jointly handle complex-background interference, weak responses to small targets, and spatial information loss under lightweight downsampling. We propose Steel-CGA, a YOLO26s-based detector that addresses these challenges for steel surface defect recognition. Steel-CGA integrates three complementary components: (1) a Channel-Spatial-Semantic Attention Module (CSSAM) appended to the backbone, which reorganizes channel, spatial, and semantic attention as three parallel branches-departing from the sequential pipeline of SE and CBAM-and couples them through residual multiplication, to suppress complex-background interference; (2) a Grouped Dilated Residual Fusion (GDRF) module redesigned from the dilation-wise residual idea of DWRSeg with an asymmetric squeeze-and-dilate channel allocation and SiLU activations, embedded in the C3k2 blocks of the neck, which enlarges the receptive field to enrich multi-scale context and strengthen responses to small defects; and (3) an Adaptive Downsampling (ADown) module adopted directly from YOLOv9, which replaces the conventional downsampling layers in the backbone, reducing spatial information loss while preserving inference efficiency. On NEU-DET and GC10-DET, Steel-CGA achieves 76.6% and 62.0% mAP@50, outperforming YOLO26s by 3.5 and 3.6 points, while reducing parameters by 26.0% and FLOPs by 23.7% and running at 134 FPS.