A Multi-Scale Sequence Fusion YOLO Model for Surface Defect Detection of Shock Absorber Connecting Rods

The shock absorber is a crucial component of the automotive suspension system, and the connecting rod, as the key part that supports its reciprocating motion, has a considerable influence on the durability and performance of the shock absorber. To address the challenges of identifying small object defects on the surface of shock absorber connecting rods, we propose a multi-scale sequence fusion YOLO model, which features a new detection layer for small objects and a restructured Neck with integrated Scale Sequence Feature Fusion (SSFF) and Triple Feature Concat (TFC) modules, significantly improving small defect detection through effective multi-scale feature extraction and fusion. Additionally, we introduce Shape-IoU as a new loss function, addressing the problem of significant label size differences in the dataset, and resulting in more accurate bounding box regression. Experimental results demonstrate that while maintaining an advantage in the number of parameters, our model achieves the highest mean average precision (mAP) of 91.5%, an improvement of 3.2% over the YOLOv8n model. For the difficult-to-identify defect categories of rottenspot and gencircle, our model respectively improved the accuracy by 12.4% and 3.9%, making it exceptionally competent at detecting surface defects on shock absorber connecting rods.