Traditional Chinese medicine decoction pieces refer to drugs that can be directly used in traditional Chinese medicine clinical or formulation production after being processed. To address the problem of difficult identification caused by the wide variety and similar morphology of traditional Chinese medicine decoction pieces, this paper proposes a classification model(ECA-Res2Net) for traditional Chinese medicine decoction pieces that integrates multi-scale features and attention. Firstly, the Res2Net residual structure is used to extract feature combinations of different quantities and scales, enhancing the network"s ability to learn detailed and semantic information; Secondly, introducing efficient channel attention to highlight the important feature regions of traditional Chinese medicine slices can improve the model"s feature selection ability; Finally, by combining Softmax Loss and Center Loss, a joint loss function is constructed to effectively adjust intra - and inter class distances and improve classification accuracy. The experiment shows that the recognition accuracy of ECA-Res2Net on the constructed dataset of 16 types of traditional Chinese medicine decoction pieces is as high as 96.35%. Compared with classical neural networks and Transformer classification models, it has fewer parameters and computational complexity, providing strong support for the fast and intelligent recognition of traditional Chinese medicine decoction pieces.