基于干預(yù)注意力的細(xì)粒度圖像識(shí)別

打開文本圖片集

中圖分類號(hào)：TP391.4 文獻(xiàn)標(biāo)志碼：A

Fine-grained image recognition based on interventional attention

CHEN Jiankun， WANG Yongxiong， PAN Zhiqun （ ， ， ， ）

Abstract： Attention plays a key role in fine-grained image recognition tasks. In order to make the model pay more atention to discriminative regions， a new method based on interventional attention was proposed to provide key clues supervising atention to learn features. Specifically， the interventional attention was added to the training process， the attention mechanism was applied to the process data cuting dropping to guide the model to improve the learning efficiency. At the same time， the fused attention was applied to the feature extraction network to help the network learn more discriminable features. In addition， label smoothing loss function center regularization loss function were introduced into the objective function， which effectively improved classification accuracy.

Experimental results show that the proposed method has excellent permance， achieving 89.8% 95.7% 94.7% classification accuracy on CUB-200-2011， St Cars FGVC Aircraft dataset respectively. In comparison with the other mainstream fine-grained classification algorithms， the proposed method achieves better classification results.

Keywords： fine-grained image recognition; interventional attention; data augmentation; fused attention; label smoothing

近幾年，細(xì)粒度圖像識(shí)別[1一直是研究熱點(diǎn)和難點(diǎn)之一。（剩余15183字）