A Comparative Study of Deep Transfer Learning Techniques for Tomato Leaf Disease Diagnosis

Abstract

AbstractThis study developed and analyzed the performance of ten different deep learning models using current transfer learning techniques to detect tomato leaf diseases. The transfer learning techniques were trained using 3000 tomato leaf images from the standard leaf disease dataset. The dataset consists of 3600 images on nine diseased and one healthy class. 300 diseased and healthy leaf images were used to validate the performance of the models. Validation performance of the transfer learning techniques was compared after 300 training epochs. The validation accuracy and loss of the DenseNet201 were 0.976 (97.6%) and 0.05476, respectively, after the 300 training epoch. Also, the test performance of the models was compared using previously unseen 300 images from the dataset. The average testing accuracy of the DenseNet201 model on tomato leaf disease detection was 96.24%. The performance of the DenseNet201 was superior to other transfer learning techniques on test data. Based on the comparison result, the training process of DenseNet201 based leaf disease detection model was extended to 1000 epochs on the dataset. The average testing accuracy of the trained DenseNet201 based tomato disease detection model was 98.31% on the test dataset.KeywordsDeep learningDenseNetLeaf disease diagnosisGraphical processing unitTransfer learning