Using BiFPN to improve YOLOv5s for empty bottle surface defect detection

With the continuous improvement of social development, food safety issues have attracted increasing attention, and the cleaning and reuse of non-disposable empty glass bottles has become a new major concern. Recycled empty glass bottles are prone to scratches, cracks and other defects due to collision and extrusion, which not only affects product quality but also endangers the health of consumers. Traditional detection methods usually rely on manual visual screening, which has low detection efficiency and cannot be guaranteed. Therefore, this paper uses deep learning technology to propose a method for detecting defects in the bodies of empty glass bottles based on the improved YOLOv5s algorithm. First, the P2 small target detection head is added to improve the network's detection accuracy for small target defects. Secondly, WIoU-Loss is used instead of the original CIoU-Loss function to significantly improve the detection ability of small targets and difficult samples. Finally, the BiFPN feature fusion structure is used to optimize FPN+PANet to improve the algorithm's recognition ability for targets of different scales. Experimental results show that the improved YOLOv5s algorithm improves [email protected] by 3.5% and [email protected]:0.95 by 3.4%. The improved algorithm shows higher accuracy and more stable robustness for target detection of defects of different scales.