Improved Image Classification with Token Fusion

In this paper, we propose a method to improve image classification performance using the fusion of CNN and transformer structure. case CNN, information about local area on an can be extracted well, but global extraction is limited. On other hand, has advantage in extraction, it requires much memory compared CNN. We apply consider feature vector each pixel resulting map by as token. At same time, divided into patches, patch considered token, like transformer. Tokens have advantages extracting information, respectively. assume that combination these two types tokens will improved characteristic, show through experiments. three methods fuse having different characteristics: (1) late token with parallel structure, (2) early (3) layer-by-layer. The proposed shows best experiments ImageNet-1K.