CFF-YOLO: cross-space feature fusion based YOLO model for screw detection in vehicle chassis

Proper installation of chassis screws is critical for vehicle quality and safety. With the widespread application of the YOLO model in the industry, we propose a Cross-space Feature Fusion based on the YOLO model for screw detection in vehicle chassis, named CFF-YOLO. We design a Cross-space Feature Fusion (CFF) module to adaptively aggregate features at different scales and correlate the low-level with high-level feature maps. According to the property of screw objects with the same scale, we modify the Yolov5 to accelerate inference speed by keeping one detection head while removing the unimportant network pathways and the other two detection heads. Besides, we design a wide-range, multi-camera line-scan imaging method to capture the same scale of screws in the whole chassis and create a custom vehicle chassis dataset (VCD). Experimental results on dataset VCD show that our proposed CFF-YOLO only takes 6.2 ms to detect one image and merely 781.2 ms to inspect an entire vehicle chassis, and outperforms Yolov5s and Yolov8n in mean Average Precision (mAP) by 6.3% and 2.1%, reaching 81.0% mAP respectively. Our proposed CFF-YOLO achieves a good trade-off between accuracy and speed.