Abstract: In order to solve the problem of large intra-class differences and high inter-class similarity in fine-grained image classification tasks, this paper proposes a multi-scale feature fusion network, which aims to learn and refine more discriminant features from the potentially different features. Among them, the feature pyramid is used as the benchmark network to obtain the difference level semantic features, and the multi-level information can provide a more comprehensive decisionmaking basis for fine-grained image classification tasks. The attention mechanism combines spatial information and channel information to enrich the convolution features. Experimental verification shows that compared with other advanced models, the proposed method effectively improves the classification accuracy, and the recognition accuracy of the vehicle dataset is 93.5%.

Key words: fine-grained image classification; vehicle identification; feature fusion; feature pyramid

摘要： 为解决细粒度图像分类任务存在类内差异性大和类间相似性高的问题，文章提出了一 种多尺度特征融合网络，旨在从具有差异性的潜在性特征中学习和提炼出更多的可判别性特 征。其中以特征金字塔为基准网络获得差异层次语义特征，多层次的信息能够为细粒度图像 分类任务提供更加全面的决策依据；注意力机制结合空间信息与通道信息，丰富卷积特征。 通过实验验证，与其他先进模型相比文章所提出的方法有效地提升了分类精度，车辆数据集 识别准确率为 93.5%。

关键词: 细粒度图像分类；车辆识别；特征融合；特征金字塔