Abstract: There are two challenges in the task of fine-grained image classification: One is that the distinguishing features are extremely subtle and difficult to capture accurately. Second, it is difficult to effectively locate the key areas of interest in the image. In this paper, a fine-grained vehicle classification recognition model based on mixed attention mechanism is proposed. By fully integrating local details and global context information, the recognition ability of fine-grained vehicle categories is significantly improved. The experimental results show that the classification accuracy of the proposed attention mechanism model on the Stanford Cars and Web-Cars open vehicle data sets reaches 94.2% and 88.0%, respectively. Compared with other network models, this model shows higher classification accuracy and stronger generalization ability. It can effectively make up for the shortcomings of traditional convolutional neural networks.

Key words: fine-grained classification; deep learning; attention mechanisms; feature fusion

摘要： 针对细粒度图像分类任务中存在的两大挑战：一是区分性特征极其细微，难以准确捕 捉；二是难以有效定位图像中感兴趣的关键区域。文章提出一种基于混合注意力机制的细粒 度车辆分类识别模型，通过充分整合局部细节信息与全局上下文信息，显著提升了对细粒度 车辆类别的辨识能力。通过实验结果表明，提出的注意力机制模型在斯坦福汽车（Stanford Cars）和网络汽车（Web-Cars）公开车辆数据集上分类准确度分别达到 94.2%和 88.0%，相较 于其他网络模型，该模型展现出更高的分类准确率和更强的泛化能力，能够有效弥补传统卷 积神经网络存在的不足。

关键词: 细粒度分类；深度学习；注意力机制；特征融合