Report of our project

Directories

analysis/ -> data analysis in jupiter notebook
app/ -> web part
  api/
  client/
data/ -> Data sources
  augmentation/ -> Data augmented
  no_augmentation/ -> Data not augmented
  models_saved/ -> Models and weights saved
microservices/
  econome/ -> Micro service which manage api instance (only for production)
preprocess/ -> scripts for preprocessing
process/ -> scripts for training (machine learning and deep learning)
utilities/ -> global utilities functions

Setup and installation

This script installs the dependencies for the project and the dataset.

./setup.sh

Preprocess dataset

This program allows you to preprocess the images, extract data and make a dataset in h5 format.

python preprocess/main.py -h

usage: python preprocess/main.py [-h] [-a] [-rmbg] [-src SRC_DIRECTORY] [-wi] [-crop] [-nor] [-nortype NORMALIZE_TYPE] [-c CLASSIFICATION] [-n NUMBER_IMG] [-rt RESULT_TYPE] [-dst DESTINATION] [-f FEATURES] [-s SIZE] [-v]

optional arguments:
  -h, --help            show this help message and exit
  -a, --augmented       Use directory augmented
  -rmbg, --remove-bg    Remove background before preprocess
  -src SRC_DIRECTORY, --src-directory SRC_DIRECTORY
                        Directory source who can find images. default (data/{augmented})
  -wi, --write-img      Write images (png) in the new directory
  -crop, --crop-img     Remove padding around leaf
  -nor, --normalize-img
                        Normalize images, you can specify the normalization type with the option -nortype
  -nortype NORMALIZE_TYPE, --normalize-type NORMALIZE_TYPE
                        Normalize images features with cv.normalize (Default: NORM_MINMAX) 
                        Types: https://vovkos.github.io/doxyrest-showcase/opencv/sphinx_rtd_theme/enum_cv_NormTypes.html
  -c CLASSIFICATION, --classification CLASSIFICATION
                        Classification type: HEALTHY_NOT_HEALTHY, ONLY_HEALTHY, NOT_HEALTHY, ALL (default)
  -n NUMBER_IMG, --number-img NUMBER_IMG
                        Number of images to use per class to select maximum of all classes use -1. (default 1000)
  -rt RESULT_TYPE, --result-type RESULT_TYPE
                        Type of result image for DP: GRAY, GABOR, CANNY, RGB. (default: GRAY)
  -dst DESTINATION, --destination DESTINATION
                        Path to save the data. (default: data/preprocess)
  -f FEATURES, --features FEATURES
                        Features to extract separate by ","
                        Example: -f=graycoprops,lpb_histogram,hue_moment
                        List of features:
                           - For DP: rgb,gray,canny,gabor,lab,hsv
                           - For ML: graycoprops,lpb_histogram,hue_moment,haralick,histogram_hsv,histogram_lab,pyfeats
  -s SIZE, --size SIZE  Size of images. (default 256x256)
  -v, --verbose         Verbose

Image augmentation

This program allows you to augment the images, and output this augmented images in the new directory.

python preprocess/image_augmentation.py -h

usage: python preprocess/image_augmentation.py [-h] [-src SRC_DIRECTORY] [-dst DST_DIRECTORY] [-ncls NUMBER_IMG_BY_CLASS] [-naug NUMBER_IMG_AUGMENTATION] [-dup] [-v]

optional arguments:
  -h, --help            show this help message and exit
  -src SRC_DIRECTORY, --src-directory SRC_DIRECTORY
                        Source of the directory where the images to augment are located. default (data/no_augmentation)
  -dst DST_DIRECTORY, --dst-directory DST_DIRECTORY
                        Source of the directory where the images to augment will be stored. default (data/tmp)
  -ncls NUMBER_IMG_BY_CLASS, --number-img-by-class NUMBER_IMG_BY_CLASS
                        Number of images to use per class to select maximum of all classes use -1. (default -1)
  -naug NUMBER_IMG_AUGMENTATION, --number-img-augmentation NUMBER_IMG_AUGMENTATION
                        Total number of augmented images, including original images. (default 3000)
  -dup, --remove-duplicate
                        Remove duplicate images created by the augmentaion process
  -v, --verbose         Verbose

Deep Learning process

This program allows to train a deep learning model with the data preprocessed

python process/deep_learning/main.py -h

usage: process/deep_learning/main.py [-h] [-p PATH_DATASET] [-lt] [-es] [-b BATCH_SIZE] [-lr LEARNING_RATE] [-opt OPTIMIZER] [-e EPOCHS] [-m MODELS] [-s] [-dst-l DEST_LOGS] [-dst-m DEST_MODELS] [-v]

optional arguments:
  -h, --help            show this help message and exit
  -p PATH_DATASET, --path-dataset PATH_DATASET
                        Path of your dataset (h5 file)
  -lt, --launch-tensorboard
                        Launch tensorboard after fitting
  -es, --early-stop     Early stop after fitting
  -b BATCH_SIZE, --batch-size BATCH_SIZE
                        Batch size
  -lr LEARNING_RATE, --learning-rate LEARNING_RATE
                        Learning rate (default 0.001)
  -opt OPTIMIZER, --optimizer OPTIMIZER
                        Optimizer (default adam). Available: dict_keys(['ADAM', 'RMSPROP', 'SGD', 'ADADELTA', 'NADAM'])
  -e EPOCHS, --epochs EPOCHS
                        Epoch
  -m MODELS, --models MODELS
                        Select model(s), if grid search is enabled, you can select multiple models separate by ",". example -m=vgg19,resnet50. By default is select all models.
                        Models availables:
                        VGG16,VGG16_PRETRAINED,VGG19,VGG19_PRETRAINED,RESNET50,RESNET50_PRETRAINED,CONVNEXTTINY,CONVNEXTTINY_PRETRAINED,CONVNEXTSMALL,CONVNEXTSMALL_PRETRAINED,CONVNEXTBASE,CONVNEXTBASE_PRETRAINED,CONVNEXTLARGE,CONVNEXTLARGE_PRETRAINED,RESNET50V2,RESNET50V2_PRETRAINED,INCEPTIONRESNETV2,INCEPTIONRESNETV2_PRETRAINED,INCEPTIONV3,INCEPTIONV3_PRETRAINED,EFFICIENTNETB0,EFFICIENTNETB0_PRETRAINED,EFFICIENTNETB7,EFFICIENTNETB7_PRETRAINED,XCEPTION,XCEPTION_PRETRAINED,CLASSIC_CNN,ALEXNET,LAB_PROCESS,LAB_INCEPTIONV3_PROCESS,HSV_PROCESS,GOOGLE/VIT-BASE-PATCH16,GOOGLE/VIT-BASE-PATCH32,GOOGLE/VIT-LARGE-PATCH16,GOOGLE/VIT-LARGE-PATCH32,FACEBOOK/CONVNEXT-BASE,FACEBOOK/CONVNEXT-LARGE,FACEBOOK/CONVNEXT-XLARGE.
  -s, --save-model      Save model
  -dst-l DEST_LOGS, --dest-logs DEST_LOGS
                        Destination for tensorboard logs. (default logs/tensorboard)
  -dst-m DEST_MODELS, --dest-models DEST_MODELS
                        Destination for model if save model is enabled
  -v, --verbose         Verbose

Deep learning models availables:

VGG

• VGG16
• VGG19

ResNet

• RESNET50

Convnext

• CONVNEXTTINY
• CONVNEXTSMALL
• CONVNEXTBASE
• CONVNEXTLARGE

ResNet & Inception & Xception

• RESNET50V2
• INCEPTIONRESNETV2
• INCEPTIONV3
• XCEPTION

EfficientNet

• EFFICIENTNETB0
• EFFICIENTNETB7

Transformers

• GOOGLE/VIT-BASE-PATCH16
• GOOGLE/VIT-BASE-PATCH32
• GOOGLE/VIT-LARGE-PATCH16
• GOOGLE/VIT-LARGE-PATCH32
• FACEBOOK/CONVNEXT-BASE
• FACEBOOK/CONVNEXT-LARGE
• FACEBOOK/CONVNEXT-XLARGE

Homemade models

• CLASSIC_CNN
• ALEXNET
• LAB_PROCESS
• LAB_INCEPTIONV3_PROCESS
• HSV_PROCESS

ML CLassification process

This program allows you to train a ML classification models with the data preprocessed

python process/ml_classification/main.py -h

usage: main.py [-h] [-cs CLASSIFICATION_STEP] -f FILENAME [-dst PROCESS_OUTPUT] [-sd] [-th THRESHOLD] [-nortype NORMALIZE_TYPE] [-cm CLASSIFICATION_MODELS] [-ct CLASSIFICATION_TYPES] [-sm]
               [-dms DEST_MODELS_SAVED] [-v]

options:
  -h, --help            show this help message and exit
  -cs CLASSIFICATION_STEP, --classification-step CLASSIFICATION_STEP
                        Classification step: LOAD_DATA, FIT_MODEL, PREDICT_MODEL, FIT_PREDICT_MODEL, ALL (default)
  -f FILENAME, --filename FILENAME
                        path and file name of the input data
  -dst PROCESS_OUTPUT, --process-output PROCESS_OUTPUT
                        Path to save or to get the preprocessed data, plots and reports. default: data/process/ml_classification
  -sd, --save-data      Save options_datasets json file and converted data from h5 format to DataFrame one with flag train/test flag, default True
  -th THRESHOLD, --threshold THRESHOLD
                        Threshold used for the filter method to select features
  -nortype NORMALIZE_TYPE, --normalize-type NORMALIZE_TYPE
                        Normalize data (NORM_STANDARSCALER or NORM_MINMAX normalization) (Default: NORM_MINMAX)
  -cm CLASSIFICATION_MODELS, --classification-models CLASSIFICATION_MODELS
                        Classification models: XGC, ETC, RFC, ALL (default). Example -cm=RFC,ETC
  -ct CLASSIFICATION_TYPES, --classification-types CLASSIFICATION_TYPES
                        Classification type: PLANTS, HEALTHY, PLANTS_DESEASES classes, ALL (default)
  -sm, --save-model     Save model, default True
  -dms DEST_MODELS_SAVED, --dest-models-saved DEST_MODELS_SAVED
                        Path to save models. default: data/models_saved
  -v, --verbose         Verbose

Machine learning models availables:

• XgBoost
• Extra tree classifier
• Random forest

Web part

With docker

cd app
docker compose up

Manualy

Start api

cd app/api
pip install -r requirements.txt
python main.py

Start client

cd app/client/
yarn
yarn dev