Steel surface defect detection is a core aspect of industrial quality control. Given the insufficient robustness of existing RGB single-modality-based defect detection models, which often suffer from high rates of false positives, false negatives, and incorrect detections of spatial morphological defects, the parallel multi-modal spatial-aware fusion YOLOv8 (PMSF-YOLOv8) algorithm was proposed. This algorithm employs a dual-branch heterogeneous network to enhance the learning of RGB texture and depth spatial features. In the mid-fusion stage, the dual-modal feature fusion module (DFFM) was utilized to achieve adaptive fusion of multiscale features through dynamic weights. The NUE-RSDDS-AUG dataset was used for validation. The results show that the PMSF-YOLOv8 network model achieves a detection accuracy of mAP@0.5 of 98.6%, with a false alarm rate reduced by 2.1% compared with that of single-modality methods, striking a balance between "high accuracy and low false alarms".