Transfer Learning based Image Classification of Diseased Tomato Leaves with Optimal Fine-Tuning combined with Heat Map Visualization

Abstract

Plant disease detection and disease classification at initial stages for sensitive commodities like tomatoes and potatoes is highly mandated as the harvest losses have a direct impact on the price fixation of the vegetables. The most identified limitation in the study of plant pathology is the availability of datasets with visual symptoms that covers all the possible diseases of one crop or plant species. Computer Vision systems and advancements in deep learning-based modeling methodologies gained significant attention in smart farming. It is presumed that the implementation of deep learning algorithms demands a large amount of data to learn complex features automatically and this can pose a challenge for applications with lesser data to achieve generalization. In such cases, Transfer Learning with optimum regularization techniques and fine-tuning mechanisms is the solution to overcome the limitations of smaller datasets. The objective of the work is to develop Tomato Disease Classification System using a transfer learning approach for ten tomato disease classes of the PlantVillage dataset downloaded from the Kaggle platform. Inception V3, a pre-trained transfer learning model is used to classify this multi-class, imbalanced, tomato plant disease based on the leaf symptoms such as dark brown lesions, concentric rings, etc. Geometrical data augmentation is used as a regularization technique to expand the size of the dataset. Significant improvement in the performance metrics is observed when the finetuning is optimum. The training accuracy and validation accuracy of the model before and after fine-tuning are 97.08%, 83.52%, and 98.19%, 95.93% respectively. The average accuracy with augmentation and optimal fine-tuning is 98%. In addition, prediction scores in terms of precision, recall, and F1-score are obtained to visualize the rate of mispredictions across the disease classes. It is observed that the misprediction rate is high across the classes early blight, late blight, and Septoria spot due to similar visual symptoms. Further, activations are used to generate an attention map in the form of Heat Maps which are included as a post-processing step before the classification of the output. Plant Leaf Disease Classification- A web application is deployed using Streamlit Python library and Ngrok services.

Keywords

• image classification
• transfer learning
• inceptionv3
• attention maps