Leaf-Diseases-Classification 🌿 (Leaffliction)

📋 Table of contents

• Description
• Setup Project
• Data Analysis
• Data Augmentation
• Image Transformations
• Classification
• Unit Tests
• Lectures

👥 Group Members:

• Abdellah Allali
• Abderrahmane Mya

---

📜 Description

This project focuses on computer vision applications related to plant leaf diseases. It encompasses tasks such as image dataset analysis, data augmentation, image transformations, and image classification to address various aspects of plant health in the context of leaf diseases.

🛠️ Setup Project

To setup the project, you need to launch the following command:

$> git clone https://github.com/aallali/Leaf-Diseases-Classification
$> cd Leaf-Diseases-Classification
$> bash ft_setup_env.sh
$> bash ft_setup_dataset.sh
$> source venv/bin/activate

note: python version used during the making of this project : 3.10.12

📊 Data analysis

The script named 000-Distribution.py is designed for extracting and analyzing an image dataset of plant leaves. It processes images from subdirectories within the provided input directory, generating both pie charts and bar charts for each plant type in the dataset. usage: there is 2 options to visualize data distribution either for all the dataset folder or just specific subfolder inside.

option1: python 000-Distribution.py ./path/to/folder/
option2: python 000-Distribution.py ./path/to/folder/subfolder

$> python 000-Distribution.py ./dataset/
('Apple_healthy', 1640)
('Apple_scab', 629)
('Apple_Black_rot', 620)
('Apple_rust', 275)
('Grape_Esca', 1382)
('Grape_spot', 1075)
('Grape_healthy', 422)
('Grape_Black_rot', 1178)

🗃️ Data Augmentation:

A second program, named 001-Augmentation.py, has been developed to balance the dataset. It employs data augmentation techniques, including rotation, projection, scaling, blur, etc., to generate six types of augmented images for each original image. usage: op1: augment all images in a given folder to given destination op2: augment a single image to "./augmented_directory" and plot it

option1: ./001-Augmentation.py ./path/to/folder/ -f="/path/to/export_location"
option2: ./001-Augmentation.py ./path/to/image

- Augment single image

$> ./001-Augmentation.py dataset/Apple/Apple_healthy/image\ \(1337\).JPG

- Balance all dataset

$> ./001-Augmentation.py dataset/ -l="augmented_directory" 
&& ./000-Distribution.py ./augmented_directory

• all classes are equivalent now

🎞️ Image Transformation:

In this section, the 002.Transformation.py program is crafted to harness the functionality of the PlantCV library. Transformations, which are processes that alter the appearance or characteristics of images, play a crucial role in the leaf classification domain. These transformations, such as Gaussian blur, ROI (Region of Interest) object identification, and object analysis, are applied directly to plant leaf images.

Transformations are essential in leaf classification as they help enhance the quality of input images and highlight relevant features. For instance, Gaussian blur smoothens the image, reducing noise and enhancing structural details. ROI object identification focuses the analysis on specific regions of interest, ensuring that only pertinent information is considered. Object analysis provides valuable insights into the characteristics of identified objects within the images.

By incorporating these transformations, the program aims to preprocess leaf images effectively, extracting key features that contribute to accurate and meaningful leaf classification. The PlantCV library serves as a powerful tool in this process, offering a range of functions for image analysis and transformation.

1. transform single given image and visualize the output

$> ./002-Transformation.py dataset/Apple/Apple_healthy/image\ \(1337\).JPG

2. transform all images in given source folder to given destination folder

$> ./002-Transformation.py augmented_datasets_train -dst="augmented_datasets_train_transformed"
Source directory :  augmented_datasets_train
Destination directory :  augmented_datasets_train_transformed
Bulk transformer is running now, please be patient...
100%|███████████████████████████████████████████████████████████| 11880/11880 [06:30<00:00, 30.42it/s]

🤖 Classification:

In the last phase, the development process includes the creation of two distinct programs: 003-train.py and 004-predict.py.

1- Train:

Within the 003-train.py program, augmented images are employed to discern the distinctive features of designated leaf diseases. This involves leveraging a Convolutional Neural Network (CNN) implemented using the Keras framework. The acquired learning outcomes are then stored and provided in the form of a compressed .zip archive.

1.1 generate config file:

to run our training model, we have to give it some params. let's generate the config.yml file first by this command:

> ./003-train.py -gc
Default configuration file generated successfully.
> ls -la config.yaml
-rw-rw-r-- 1 allali allali 99 Feb  6 16:44 config.yaml

config.yml:

epochs: 5
model: path/to/existing/trained/model.h5
model_save_location: models
training_set: /path/to/augmented_datasets_train_transformed

explanation: epochs : number of times to train the model model : if you have already trained a model and want to train over it again (set it to blank to start fresh) model_save_location : location where the final trained model will be saved (default name model.h5) training_set : path to dataset location to train over

1.2 train your model:

after setting your config.yml start the trainin with this command:

> ./003-train.py
Found 79210 files belonging to 8 classes.
Classes saved: ['Apple_Black_rot', 'Apple_healthy', 'Apple_rust', 'Apple_scab', 'Grape_Black_rot', 'Grape_Esca', 'Grape_healthy', 'Grape_spot']
Epoch 1/10
 205/1980 [==>...........................] - ETA: 5:28 - loss: 1.5214 - accuracy: 0.4284

to reduce the hardware stressing, (like a laptop or low end PC), we stop the training for 60 seconds after every 10 epoches done, this will give some time for the hardware to take a rest and reduce the temperature 🔥 a little bit before start again, especially if you training the model on a GPU.

+ at each pause (60s) we save the model for the current epoch 💾 reached, to keep a copy of the model in case something happend at the end, you find yourself with most recent model aquired. eg: model_progress_10.h5, model_progress_20.h5, ... at the end 2 files will b generated :

• labels.txt : containg the calsses names
• model.h5 : containg the calsses names
• model_progress_{step}.h5 : containg the calsses names

2- Prediction:

On the other hand, the 004-predict.py program is designed to take a leaf image as its input. It not only displays the original image but also showcases its various transformations. Furthermore, the program makes predictions regarding the specific type of disease present in the given leaf.

2.1 Predict single leaf:

./004-predict.py dataset/Apple/Apple_healthy/image\ \(1337\).JPG

2.2 Predict batch of leafs:

to test the accuracy of our model, we added another option to test over a batch of max 100 random image from a given folder recursively. you can use this command to run the prediction over our datasets. $> clear && ./004-predict.py /path/to/folder -m model/model.h5

result of 8 random 100 images in dataset/Apple folder:

92/100 (92.0%) predicted correctly
91/100 (91.0%) predicted correctly
87/100 (87.0%) predicted correctly
91/100 (91.0%) predicted correctly
96/100 (96.0%) predicted correctly
93/100 (93.0%) predicted correctly
92/100 (92.0%) predicted correctly
87/100 (87.0%) predicted correctly

result of 8 random 100 images in dataset/Grape folder:

88/100 (88.0%) predicted correctly
86/100 (86.0%) predicted correctly
86/100 (86.0%) predicted correctly
86/100 (86.0%) predicted correctly
86/100 (86.0%) predicted correctly
86/100 (86.0%) predicted correctly
85/100 (85.0%) predicted correctly
87/100 (87.0%) predicted correctly

$ ./004-predict.py -h
usage: 004-predict.py [-h] [-lb LABELS] [-m MODEL] image_path

Predict class of a leaf image or directory

positional arguments:
  image_path            Path to the image

options:
  -h, --help            show this help message and exit
  -lb LABELS, --labels LABELS
                        /path/to/labels.txt (default: models/labels.txt)
  -m MODEL, --model MODEL
                        /path/to/model.h5 (default: models/model.h5)

3- final step:

this step is not necessary for the concept realisation of CNN, but its required by the subject. after we trained our model let compress what we used to build/train our model in a zip file

• created a folder and moved my dataset + labels + model

> ls model
dataset  labels.txt  model.h5
> zip -r model.zip model
> ls -la model.zip
-rw-rw-r-- 1 xxxxx xxxxx 184552738 Feb  6 15:38 model.zip

---

🔬 Unit Tests:

we tried to write as much unit tests as possible for the program especially the helpers functions (libft)

• run all tests:

python3 -m unittest discover -s ./unittests -t .. -v

• run single test file:

python3 -m unittest unittests/libft/test_ft_dict.py -v

---

📖 Lectures:

Check docs folder for lectures about classification