Developing YOLOv5s model with enhancement mechanisms for precision parts with irregular shapes

Abstract

Given the precision issues caused by part irregularities and occlusion when intelligent assembly technology is used to identify and classify aerospace parts, this study develops a You Only Look Once Version 5 Small (YOLOv5s) model with enhancement mechanisms for detecting precision parts with irregular shapes. First, offsets and learnable parameters are employed in the convolutional layer and combined with the Cross Stage Partial Bottleneck with 3 convolutions (C3)module of the YOLOv5s neck network to enhance the model’s feature extraction capability for irregular objects. Second, the processing speed and recognition accuracy are increased by optimizing the loss function using the smallest possible distance between the corner points of the predicted boxes and the ground truth. Finally, a method is proposed for converting spatial information to channel information in the backbone and neck, thereby reducing information loss and enhancing detection accuracy for small targets. In this study, a dataset of precision parts with irregular shapes based on aerospace components was conducted and experimentally validated. The findings show that the suggested method outperforms YOLOv5 and the most current YOLOv9, increasing identification accuracy to 93.7 % and achieving a speed of 102.04 frames per second. This approach delivers improved detection accuracy for tiny targets, occlusions, and irregularly shaped components as compared to two-stage and one-stage detection algorithms. Furthermore, it maintains a pace of 100 frames per second while striking an optimal balance between speed and accuracy, offering a practical solution for the quick and accurate