Improvement of Vision Transformer Using Word Patches

Ayato Takama, Sota Kato, Satoshi Kamiya, Kazuhiro Hotta

Abstract: Vision Transformer achieves higher accuracy on image classification than conventional convolutional neural networks. However, Vision Transformer requires more training images than conventional neural networks. Since there is no clear concept of words in images, we created Visual Words by cropping training images and clustering them using K-means like bag-of-visual words, and incorporated them into Vision Transformer as ”Word Patches” to improve the accuracy. We also try trainable words instead of visual words by clustering. Experiments were conducted to confirm the effectiveness of the proposed method. When Word Patches are trainable parameters, the accuracy was much improved from 84.16% to 87.35% on the Food101 dataset.