prvnsingh/plantdiseaserecongnition

Comparison of various CNN models on different datasets

PlantDiseaseRecongnition

The project aimed at developing an automated plant disease classification system that can assist farmers, researchers, and other stakeholders in the agricultural sector. The project leverages digital imaging technology and deep learning algorithms to accurately detect and classify different types of plant diseases using leaf images.

We have used deep learning algorithms such as MobileNet, ShuffleNet, and Resnet18 to learn the visual features of healthy and diseased leaves and classify them into different disease categories.

Specifically, the models achieved accuracy rates ranging from 92.64% to 99.19% on various datasets of diseased and healthy plant leaves. These results demonstrate the potential of deep learning-based solutions for accurately identifying and classifying plant diseases, which could ultimately reduce the need for manual inspection by experts and improve the efficiency and cost-effectiveness of plant disease diagnosis.

Getting started

We have trained 3 CNN models

• MobileNet-v2
• ResNet18
• ShuffleNet

In each of above directory we have three subdirectory containing jupyter notebook for each Dataset.

• Dataset 1
• Dataset 2
• Dataset 3

Instruction to obtain the Datasets: Dataset can be downloaded directly from the Kaggle. In the included notebooks, we just have to define the Username and the password for Kaggle. And the dataset can be downloaded through the implementation it self.

Eg:

import os
os.environ['KAGGLE_USERNAME'] ='sampleUserName'
os.environ['KAGGLE_KEY'] = 'samplePassword'

By running the below code in the implementation notebooks, the dataset will be downlaoded to the colab environment.

  !kaggle datasets download -d rashikrahmanpritom/plant-disease-recognition-dataset

Floder Structure:

HyperParameterTuning

Contains the notebooks which is used for the hyper parameter tuning.
  Dataset1
    There several notebooks which uses the dataset 1 with shufflenet.
      shufflenet1 _epoch10_batch62.ipynb
      shufflenet1 _epoch20_batch32.ipynb
      shufflenet1_epoch10_batch128.ipynb
      shufflenet1_epoch_1.ipynb
      shufflenet1_epoch_20_batch62.ipynb
  Dataset2
    There 4 notebooks which uses the dataset 2 with shufflenet for the batch sizes 32, 64, 128 and 256.
      shufflenet2new_batch_size_128.ipynb
      shufflenet2new_batch_size_256.ipynb
      shufflenet2new_batch_size_32.ipynb
      shufflenet2new_batch_size_64.ipynb

MobileNet-v2

There are 3 models of MobileNet-v2 with 3 datasets
  Model1/MobileNetv2_dataset1.ipynb
  Model2/MobileNetv2_dataset2.ipynb
  Model3/MobileNetv2_dataset3.ipynb

ResNet18

There are 3 models of ResNet18 with 3 datasets and TSNE model along with the saved models.
  Dataset1
    dataset-1.ipynb
    Model.pt
    Model.pth
  Dataset2
    Dataset-2.ipynb
    Model.pt
  Dataset3
    Final_Dataset3.ipynb
    Model.pt
  TSNE
    TSNE_ResNet18_Dataset3.ipynb

ShuffleNet

There are 3 models of ShuffleNet with 3 datasets and TSNE model along with the saved models.
  ShuffleNet_Dataset_1
    model.pt
    model.pth
    shufflenet1.ipynb
  ShuffleNet_Dataset_2
    model.pt
    shufflenet2.ipynb
  ShuffleNet_Dataset_3
    model.pt
    shufflenet3.ipynb

TSNE

There are 3 TSNE models using 3 datasets along with MobileNet-v2 and ShuffleNet
  Dataset1
    TSNEMobileNetv2_model1.ipynb
    TSNEshufflenet1.ipynb
  Dataset2
    TSNE_MobileNetv2_model2.ipynb
    TSNE_shufflenet2.ipynb
  Dataset3
    TSNE_shufflenet2.ipynb

TransferLearning

There are 2 models with Transfer learning using MobileNet-v2 and ShuffleNet
Model1
  MobileNetv2_model2.ipynb
Model2_shuffleNet
  best_model.pt
  transferlearning-shufflenetv2.ipynb

Requirements: The following libraries have been used for the implementation of these 9 models.

* os
* shutil
* random
* torch
* torchvision.transforms
* torchvision.datasets
* torch.nn
* torch.optim
* torchvision.models
* ImageFolder
* glob
* matplotlib.pyplot
* tqdm
* numpy
* sklearn.metrics
* seaborn

Hyperparameter Tuning

The directory HyperParameterTuning contains different notebook run on different hyper-parameters.

TNSE

The directory TNSE contains notebook showing TNSE analysis of datasets

Transfer Learning

The directory transfer learning contains notebooks in which model are run on pretained weights such as Weights.IMAGENET1K_V1

How to run

Import the notebook at google collab and run in GPU.

Testing

Each model directory has saved trained model .pt file. You can import the saved model and do inference on your custom dataset to detect plant disease.