Heterogeneous domain adaptation method for tomato leaf disease classification base on CycleGAN

Abstract

Despite the significant improvements in the detection and diagnosis of plant diseases at an early stage facilitated by deep learning technology, there are challenges associated with the generalization performance of deep learning models. These problems from the differences between in-field and in-lab data, as well as the heterogeneity of training and prediction data features. In the case of tomato leaf diseases, the PlantVillage dataset is widely used and has already demonstrated accuracy of more than 99%. However, using trained model based on this dataset to predict in-field data results in low accuracy due to domain differences and heterogeneous features. In this paper, we propose a domain adaptation method based on CycleGAN to solve this problem, followed by a preprocessing technique that utilizes both the OpenCV module and a segmentation model based on U-Net for the best generalization performance. The classification accuracy is evaluated by applying the DenseNet121 model trained on the PlantVillage dataset to the images generated by CycleGAN. Our results demonstrate, with an F1-score of 95.6%, that our domain adaptation method between the two domains is effective in mitigating the effect of domain shift.