Transfer Learning Based Benchmarking of Convolutional Neural Network Architectures for Tomato Leaf Disease Classification
Abstract
Detection of tomato foliage diseases at an initial and accurate stage is crucial for increasing productivity and minimizing agricultural losses. Deep learning has proved its effectiveness in automated plant diseases diagnosis; yet the problem of choosing the suitable CNN architecture remains topical. In this paper, we conduct comparative benchmarking analysis of five CNN architectures, such as MobileNetV3, EfficientNet-B0, ResNet50, DenseNet121, and ConvNeXt-Tiny, that have proven their effectiveness in other applications for automated classification of tomato foliage diseases through transfer learning. To attain this goal, a widely available tomato foliage disease dataset is used where the original training subset is divided into training and validation subgroups, and the provided validation subset is taken as the independent test set. Preprocessing and augmentation techniques are used for image processing, and the weighted cross-entropy loss with label levelling is chosen to report the problem of class inequality during the training process. All five CNN models were trained under the same experimental setting with pretrained ImageNet weights and frozen feature extractors. Accuracy, precision, recall, F1 score, confusion matrix analysis, and multi-class ROC-AUC were chosen as evaluation metrics. The comparative analysis results offer important lessons regarding the classification abilities of various CNN models and form the basis for future research work on automated disease classification of tomato leaves.