A Fine-Grained Car Recognition Method Based on a Lightweight Attention Network and Regularized Fine-Tuning

Abstract

Car fine recognition is a typical scenario for fine-grained image classification, which has great research and application value in both civilian and military fields. However, current research on fine-grained classification is often limited to improving the accuracy of classification models, ignoring the need for lightweight and efficient applications in practical applications, resulting in a disconnect from reality. In this paper, a fine-grained car recognition method based on a lightweight attention network and regularized fine-tuning is proposed. Based on the high-performance, lightweight convolutional neural network (CNN) architecture MobileNet V3, an improved CNN architecture HAM-MobileNet that includes a hybrid attention module is designed. A regularized fine-tuning strategy that includes correlation constraints is adopted. By fine-tuning the HAM-MobileNet, accurate classification of car images can be achieved. The experimental results on the Stanford cars dataset show that the proposed method achieves an accuracy rate of 84.6%, which is the highest level among all lightweight CNN architectures and is comparable to non-lightweight CNN architectures. The visualization results show that the proposed hybrid attention module can make the network model focus more on the target objects with consistent classes, suppress task-irrelevant backgrounds and other noise, and improve the learning ability and generalization of the network model.