关键词: 细粒度图像分类, 注意力机制, 数据增强, 多粒度特征融合

Abstract: Aiming at the problems that it was difficult to accurately identify the key information of fine-grained images, the classification index was relatively simple and the feature utilization was not sufficient in existing models, we  proposed a new  fine-grained image classification network model. In the network training step, the model embedded a dual attention network to strengthen the correlation between middle-level features and depth features. According to the different receptive field sizes of different layers of the network, the data were trimmed and then spliced into new sample data as the input for the next layer. The support vector machine classifier was used to take the output results of middle-level features and depth features together as the final classification index.  The experimental results  on three classic datasets CUB-200-2011, Stanford Cars and 102 Category Flower show that the classification accuracy reaches 89.56%, 95.00% and 96.05%, respectively. Compared with other network models, it has better classification accuracy and generalization ability.

Key words: fine-grained image classification, attention mechanism, data augmentation, multi granularity feature fusion