Leaf Disease Detection

Motive

The purpose of this project is to create an application that can detect various kinds of leaf diseases. This app can detect the name of the leaf. It can also detect if the leaf is diseased or not and will also tell the disease name. It can detect seven kinds of leaves and their diseases. The leaves are - Banana, Cucumber, Mango, Maple, Pepper, Rose, Tomato. The diseases are - Sigatoka, Powdery-mildew, Anthracnose, Leaf-spot.

Data Collection

I collected my data manually. I tried to collect data automatically with the help of tools like selenium but it was not giving me accurate images. So I choose to collect data manually by Google search.

Data Preprocessing and Image Augmentation

Healthy images were available on the internet but there were not many images available of the diseased leaves. As my dataset is small I had to augment images for increasing the number of images.

Number of Images after augmentation :

Leaf Name	Total	Healthy	Diseased
Maple	1216	448	768
Banana	704	472	232
Rose	2432	1488	944
Mango	1944	1248	696
Cucumber	768	416	352
Tomato	1872	1336	536
Pepper	784	456	328

Model Training

For better performance and time effect I choose 'Fastai' and 'Pytorch'. For training, I tried two models resnet-34 and resnet-18, and saved the best-performing model for future use. Both models gave 99% accuracy. restenet-18 was faster than restnet-34 so I selected restnet-18.

Final Model Selection

Trained 4 models. Here are their results.

Model	Accuracy	F1_score	Precision	Time
Resnet18	0.998834	0.928576	0.930272	01:38
Resnet34	0.999863	0.932688	0.932644	02:11
AlexNet	0.979863	0.913688	0.912644	03:15
GoogleNet	0.999108	0.930197	0.930705	01:39

As resnet34 and resnet18 both reach 99% Accuracy, another score is way close, but resnet18 is a bit faster. So I chose resnet18 for further work.

Model size Compression

Model compression is very essential. By compressing the models with ONNX, you can achieve smaller model sizes, faster inference times, and reduced memory requirements. Model compression code can be found in the notebook /leaf-disease-detection.ipynb file.

Deployment

I deployed my project in HuggingFace. You can see this from here.

Interface

I also deployed the app using flask on render.com. You will find the code in the flask-deployment branch. The UI of the application is simple. Here is the link