Car-YOLO: an efficient multiscale feature fusion network for automotive steering bellows defect detection

Automotive steering bellows are critical dust-proof components that prevent impurities from entering the steering gear, making their defect detection and clamp positioning accuracy essential for vehicle safety. Traditional manual inspection methods suffer from low efficiency, poor consistency, and high miss rates, particularly for subtle defects like cracks and minor clamp displacements. This paper proposes Car-YOLO, an improved lightweight detection model based on YOLOv11, specifically designed for automotive steering bellows defect identification and clamp localization. First, we construct a high-quality dataset comprising 13,020 images with five key categories: break, depression, clip, locknut, and without clip. Second, we introduce three novel enhancements: a MambaC3K2 module in the backbone that integrates MambaVision with C3K2 to enhance long-sequence dependency modeling for complex bellow textures; a Global-Local Spatial Attention (GLSA) module embedded in a Bidirectional Feature Pyramid Network (BiFPN) to simultaneously capture global contextual relationships and local fine details; and a Localization Quality Estimation (LQE) head that leverages bounding box distribution statistics to ensure consistency between classification confidence and localization accuracy. Additionally, we develop an AFGC-attention enhanced R3GAN to address class imbalance by synthesizing high-quality minority defect samples. Experimental results demonstrate that Car-YOLO achieves 94.9% precision, 95.0% recall, and 97.2% mAP@0.5 with only 2.83M parameters and 310.2 FPS inference speed, significantly outperforming state-of-the-art detectors including YOLOv8, YOLOv10, YOLOv12, and Mask R-CNN.