SCSAF-YOLO: An Enhanced YOLOv5-Based Object Detection Method for Unordered Industrial Workpieces

To address the issues of degraded detection accuracy and high false-positive rates caused by densely distributed workpieces and complex background interference in industrial sorting scenarios, this paper proposes an improved object detection algorithm based on YOLOv5l, termed SCSAF-YOLO. A spatial-channel synergistic attention (SCSA) mechanism is introduced to jointly model channel dependencies and spatial localization cues, thereby enhancing feature representation under occlusion and dense object distributions. To alleviate sample imbalance and improve bounding box regression accuracy, a Focal and Efficient Intersection over Union (F-EIOU) loss is employed to replace the conventional CIOU loss. Furthermore, depthwise separable convolutions (DWConv) are partially integrated to reduce computational complexity and model parameters while preserving detection performance. Extensive experiments on a self-built industrial sorting dataset featuring multi-scene occlusion and dense object layouts demonstrate that the proposed method achieves 97.6% mAP50 and 69.2% mAP50–95, outperforming the baseline by 1.4% and 2.4%, respectively, and surpassing mainstream YOLOv8 and YOLOv10 models. Experiments on the VOC2007 dataset further verify the generalization ability, with improvements of 7.8% and 5.7% in mAP50 and mAP50–95, respectively. Ablation studies confirm the synergistic effectiveness of the proposed components, indicating the potential of the method for industrial visual sorting applications.