YOLOv5-LC: A Lightweight Classification Deep Learning Detector for Dining Plate

Plate detection refers to the process of detecting and recognizing the type of plate. This technology has a wide range of applications in fields like catering, smart homes, and healthcare. Current there are many challenges in plate detection methods based on deep learning, such as long training time, high misidentification rate, and inability to deploy on devices. To solve this problem, we propose a lightweight plate classification and pricing algorithm called YOLOv5-LC. Our algorithm (1) has a smaller model size and faster inference speed, (2) achieves accurate plate detection and classification without sacrificing accuracy, and (3) is easier to encapsulate and de-ploy on end devices. First, we improve the anchor box generation process of YOLOv5 and prune and distill the network. Second, we use an improved focal loss as the classification loss function to address the issue of difficulty in distinguishing similar samples. We also introduce Alpha-IoU to improve the convergence speed and accuracy of the algorithm. To evaluate the detection capability of the model, we conducted experiments on an NVIDIA RTX 3060Ti GPU, and the results showed that YOLOv5-LC can achieve a model volume of 1.7M (9.2% of YOLOv5) and better detection accuracy (2.3% mAP increase).