A Two-Stage Transfer Learning Method for Tea Disease Identification

Tea has been widely planted across the world, while the identification of tea disease is mainly based on human expertise. In recent years, there has been an increasing use of deep learning techniques for the recognition of plant diseases. However, persistent challenges such as the limited availability of comprehensive datasets and the need to improve recognition accuracy still exist in tea disease identification. To address these issues, we collect a small-scale tea disease image dataset, containing 203 images covering six common tea disease. Then, we proposed a two-stage transfer learning approach for improving the accuracy. Firstly, we modified the first stage in ResNet50 to better extracting image feature. Next, we leverage a two-stage transfer learning approach which first transfers to a related intermediate PlantVillage dataset. The proposed model achieved a recognition accuracy of 87.67% when trained from scratch on the tea leaf dataset, which is 5.46% higher than the ResNet50 model. Furthermore, the two-stage transfer algorithm can further improve the recognition accuracy to 90.64%.