关键词: 多肉植物, 图像分类, 对比学习, ConvNeXt, 投影层

Abstract: In view of the large variety of succulents，the large intra⁃class differences and the small inter⁃class differences，as well as the difficulty of data collection，traditional classification algorithms cannot effectively solve the problem of succulent plant image classification.This paper proposed a contrastive learning based succulent plant image classification network CL_ConvNeXt.The network was based on ConvNeXt structure and introduced the idea of contrastive learning.A non⁃linear projection layer（Projection head）was added in the middle layer of the network as an auxiliary classifier to help the model extract features from the shallow network.In a batch，positive samples were constructed through data augmentation，and the remaining samples were considered as negative samples.The cross entropy loss function and the contrastive loss function were weighted to newly design loss function calculation method，which could achieve one⁃stage model training.Transfer learning was used during training to transfer the pre⁃trained weights to the model to improve the convergence speed of the model，and various strategies and parameters were optimized to further improve the recognition accuracy of the model.The experimental results showed that on the self⁃made 190⁃class succulent plant dataset，under the same training strategy and environment configuration，the recognition accuracy of the final model CL_ConvNeXt for succulent plant image classification reached 91.79%，which was 12.24 percentage points higher than that of the original ConvNeXt model structure，showing good effect on solving the problem of succulent plant image classification and recognition.

Key words: Succulents, Image classification, Contrastive learning, ConvNeXt, Projection head