Our aim is to classify the given rice leaf image into one of the given four categories.
For comparison, we use the results from the paper Sethy et al.(link) with their best model of using ResNet for feature extraction and SVM on top of it for classification. Our models include Inception+attention, ResNet + Attention and VGG + Attention on the same dataset
| Method | ResNet + SVM (original paper) | Inception + Attention | ResNet + Attention | VGG + Attention |
|---|---|---|---|---|
| Metric | F1 score | F1 score | F1 score | F1 score |
| Bacterial blight | 98.38 | 97.27 | 96.76 | 96.35 |
| Blast | 96.43 | 96.26 | 95.75 | 94.48 |
| Brown spot | 96.70 | 97.67 | 97.55 | 95.39 |
| Tungros | 100 | 99.71 | 99.60 | 99.62 |
We use the rice leaf dataset obtained from Sethy et al. Dataset Dataset consists of 4 classes {Bacterial blight, Blast, Brown spot, Tungro.}
pytorch >= 1.2.0
pandas
numpy
tqdm
scikit-learn
torchvision
To install and use with GPU, cuda toolkit along with drivers need to be installed. And set use_cuda = 1 in training/testing codes.
This is for training 30 times with random sampling of data, where each of the 30 runs will be run for 50 epochs. Testing is done for every run once and the models and test results are stored in the same directory
git clone https://github.com/sraone-96/Rice-Leaf-Disease-Detection.git
cd Rice-Leaf-Disease-Detection
python3 train_test_all_models.py
NOTE : Make sure to have the corresponding image data in the same folder where you are running the code from. And name the folder as Rice Leaf Disease Images. For multi-head attention, change number of heads in the specific model code (eg: VGG_Model.py). Change the model to run in the code by uncommenting the required model.