Research on Recognition Method of Chinese Cabbage Growth Periods Based on Swin Transformer and Transfer Learning

Highlights To the best of our knowledge, this study was the first intelligent recognition for Chinese cabbage growth period and proposed the Swin Transformer+1 model. If the four growth periods were considered, the recognition accuracy rate of the model on the test set was 96.15%. If the transition periods of Chinese cabbage growth were considered, the model recognition accuracy rate was 97.17%. Experiments showed that the Swin Transformer+1 model was robust and could be applied in real agricultural production. Abstract. In order to facilitate agricultural management and improve the quality and yield of Chinese cabbage, it is necessary to intelligently identify the growth periods of Chinese cabbage. In this study, a transfer learning-based recognition model for Chinese cabbage growth periods was proposed, which could identify four growth periods of Chinese cabbage: “germination and seedling period,” “rosette period,” “heading period,” and “dormant period.” The data set of Chinese cabbage growth periods was built. The recognition model was named Swin Transformer+1, using Swin Transformer as the backbone network to extract image features, and a fully connected layer as the classifier. To optimize the model, we used Letterbox instead of Stretching to resize the image, used Focal Loss instead of Cross Entropy Loss as the loss function, and used Stochastic Weight Averaging instead of Adam as the optimizer. Transfer learning was used for training, which could solve the problems of overfitting and underfitting when training deep network with a small data set. We verified the effectiveness of the above improved methods through ablation experiments. Experiments showed that the Swin Transformer+1 model had a high recognition accuracy rate. If only the four growth periods were considered, the recognition accuracy rate was 96.15%. If the transition periods between two growth periods of Chinese cabbage were considered, the recognition accuracy rate was 97.17%. The model had strong robustness. It maintained a high recognition accuracy rate when the images in the test set were augmented. In general, Swin Transformer+1 model has high application value in actual agricultural production scenarios. Keywords: Chinese cabbage growth period, Deep learning, Image recognition, Swin transformer, Transfer learning