Swin-Transformer Based Classification for Rice Diseases Recognition

Abstract

Rice is easy to be infected with various diseases and pests during the growth process, resulting in the decline of rice yield. Therefore, large scale and automatic identification of diseases and pests is one of the important problems to be solved in the development of agriculture. The traditional identification of agricultural diseases and pests mainly depends on visual observation or machine learning. With the emergence of deep learning, it effectively solves the important problem of automatic identification and diagnosis of rice diseases. In this paper, we adopted a new rice disease identification method based on swin-transformer including sliding window operation, with hierarchical design. Taking pictures of rice diseases in field environment as the research object, several common rice diseases were recognized. The proposed swin-transformer model achieves an accuracy of 93.4%, and this accuracy is higher than conventional machine learning model 4.1%. Experiments show that the algorithm can effectively improve the accuracy of rice disease detection compared with traditional machine learning, and has a certain reference for other agricultural products.