Real-Time Ultra-Fine-Grained Image Recognition Using YOLOv5

Due to the similarity between classes and the difference within classes in fine-grained image recognition, it is difficult for traditional models to capture enough feature information to distinguish objects of different subclasses. This study is based on the YOLOv5 algorithm, introducing attention mechanism and Feature Pyramid Network (FPN) on the basis of YOLOv5, while increasing data diversity through image flipping, rotation, cropping, and scaling, and introduces the Res2Net CBAM module in YOLOv5 to enhance sensitivity to important information. The FPN structure is used to achieve the fusion of multi-scale features, enhance the channel self-regulation and spatial attention of features, and improve the model’s capture of key information. The CIoU loss function is used to replace the traditional IoU loss, and the Adam optimizer is selected according to the size and characteristics of the dataset to achieve more efficient training. The confidence threshold optimization strategy is applied in NMS to improve the accuracy of detection. Based on the above content, it was finally found that the improved YOLOv5 model has significantly improved compared to the traditional YOLOv5 in terms of recognition accuracy, recognition delay and missed detection rate. Specifically, the average recognition accuracy increased by 6.5%, the average recognition delay was reduced by 156.9ms, and the missed detection rate was also reduced, showing better performance stability and significantly promoting ultra-fine-grained image recognition.