Automatic Blight Disease Detection in Potato and Tomato Plants

Tools Used: Python, Pytorch, Optuna

Description
In this project, a ResNet-9 model was built and used for image classification of potato and tomato plant leaf images in order to detect blight diseases (early blight and late blight) in these images. The motivation for doing this was to come up with a tool which helps farmers detect blight diseases faster and more effectively.

Processes followed
1. Data Augmentation (center cropping, gaussian blurring and random rotation) on plant leaf images
2. Hyperparameter tuning in Optuna
3. ResNet-9 model implementation and training in Pytorch
4. VGG-16 baseline model implementation and training
5. ResNet-9 model evaluation and comparison with VGG-16 baseline model
6. Shapley values and saliency maps used to explain ResNet-9 model predictions

Results
ResNet-9 test accuracy : 99.25%
VGG-16 test accuracy : 97.90
Saliency maps proved that the ResNet-9 model works as expected by looking at diseased areas and general green areas of the leaves to make predictions


Github link