This project implements a Convolutional Neural Network (CNN) to classify plant leaf images into different categories of diseased and healthy plants. The main goal is to design and train our custom CNN model that achieves high accuracy in image classification tasks.
Utilize Convolutional Neural Networks (CNNs) to accurately identify plant diseases from leaf images.
CNNs have revolutionized image classification tasks, and this project aims to explore their potential in a real-world dataset. By the end of this project, we aim to create a model that can classify images with high accuracy, showcasing the power of deep learning.
Dataset contains 70k plant images in train set and 18k test images
14 crop diseases : Apple,Blueberry,Cherry,Corn Grape,Orange,Peach,Pepper,Potato,Raspberry,Soybean,Squash,Strawberry,Tomato
Images of 17 fungal diseases,4 bacterial diseases,2 mold diseases,2 viral diseases and 1 disease caused by mites
- Classes: 38
- Training Images: 70,000
- Test Images: 18,000
- Image Size: 256x256 pixels, RGB
- Source: PlantVillage Dataset
To prepare the dataset for training, we:
- Normalize the pixel values
- Perform data augmentation techniques such as veritcal and horizontal flip,color jitter to improve model generalization.
The CNN model used in this project consists of the following layers:
| Layer Type | Output Shape | Details |
|---|---|---|
| Conv2D + ReLU | (128, 128, 32) | Kernel size: 3x3, Padding: 1 |
| MaxPooling | (64, 64, 32) | Pool size: 2x2 |
| Conv2D + ReLU | (64, 64, 64) | Kernel size: 3x3, Padding: 1 |
| MaxPooling | (32, 32, 64) | Pool size: 2x2 |
| Conv2D + ReLU | (32, 32, 128) | Kernel size: 3x3, Padding: 1 |
| MaxPooling | (16, 16, 128) | Pool size: 2x2 |
| Conv2D + ReLU | (16, 16, 256) | Kernel size: 3x3, Padding: 1 |
| MaxPooling | (8, 8, 256) | Pool size: 2x2 |
| Flatten | (8 * 8 * 256) | Flattened output for FC layers |
| Fully Connected | (1024) | Dense layer with ReLU activation |
| Dropout | (1024) | Dropout rate: 0.2 |
| Fully Connected | (#Classes) | Output layer with softmax |
- Dropout is used after the fully connected layers to prevent overfitting.
- Batch Normalization is applied to improve convergence.
- Optimizer: AdamW
- Learning Rate: 0.0001
- Loss Function: Cross-EntropyLoss
- Batch Size: 128
- Epochs: 15
The model was trained using the Pytorch framework.
To prevent overfitting and improve generalization, we applied the following augmentations:
- Horizontal flip
- Vertical flip
- Color Jitter
Training was done on a NVIDIA T4(kaggle) .
After training the model, we evaluated it using the test set. The following metrics were calculated:
- Accuracy: 98.57%
- Precision: 0.986
- Recall: 0.9857
- F1-Score: 0.9857
Below is the confusion matrix showing the number of correct and incorrect predictions for each class:
The trained model achieves an accuracy of 98.6% on the test set. Below are some sample predictions:
Below are all the experiments i performed while training the model
These results demonstrate that the model performs well for many classes but occasionally struggles with similar-looking categories.
This project is licensed under the MIT License