Research on the Improved YOLOv5 Algorithm for Recognizing Cylindrical Parts

In response to the low performance of component recognition algorithms in the complex environment of guiding roller production workshops, a YOLOv5-SE target recognition algorithm based on attention mechanism is proposed. Firstly, the algorithm is based on the YOLOv5 network structure and, using a self-built guiding roller component dataset, data augmentation is performed through methods such as Gaussian blur and contrast enhancement. Secondly, a channel attention mechanism is added to the network's Backbone to learn weights for each channel, improving the network's feature extraction capability. Finally, by replacing CIoU Loss with Shape-IoU Loss to improve the classifier's loss function, the algorithm enhances boundary box regression rates while improving target localization accuracy and model convergence speed. Experimental results show that the detection accuracy of each component in the improved algorithm is 89.3%, which is 3.7% higher in average precision (mAP) compared to the previous algorithm.