Clothing classification based on channel attention and token labeling

With the rapid development of e-commerce platforms, people need to classify massive clothing in a more detailed manner.However,existing models like CNN-based methods are not good at capturing long-distance dependencies because of limited perceptual fields while transformer-based methods lack strenth of capturing local feature details. In this paper, we propose a ECA-TL model that connects CNN to ViT models to get both global and local features for clothing classification. Firstly, we use ResNet to extract low-level semantic information and a light weight Efficient Channel Attention module(ECA) to obtain features with channel attention.Then the self-attention mechanism of ViT model is applied to capture global features while the training process is supervised by pre-generated local labels from the TokenLabeling module(TL). The experimental results show that our proposed method has a effective improvement on the prediction ability of the benchmark network with an accuracy of 87.78%.