Skip to content

MasquitoWasTaken/masquito-ml-training

Repository files navigation

masquito-mask-detection

overview

Machine learning part of a fullstack application to tell if you're wearing a mask right

Followed ImageAI's prediction model

Used images from Kaggle:

run this program

  1. Clone this repository
  2. Install the pip dependencies listed below
  3. Download the latest model from the releases tab
  4. Place the .h5 file in training_data/models/
  5. Edit test_model.py's model variable to match the filename
  6. Place a test image (or use an included one) in test_images/
  7. Edit test_model.py's test_image variable to match the filename
  8. Run the code with python ./test_model.py

dataset construction

We used two publicly available datasets from the website Kaggle. Two scripts (mmds_to_cropped.py and fmdds_to_cropped.py) cropped out the faces labelled in each dataset. check_images.py made sure there was no corrupt data. Then we combined the processed images for both datasets into aggregate/. Finally, we manually cut down each class into 270 images (limited in quantity by the improper class) and used 220:50 train:test ratio, formatted in training_data/ using ImageAI's directory structure.

scripts

mmds_to_imageai.py and fmmds_to_cropped.py

Converts the dataset from Kaggle's format to ImageAI's format

check_images.py

For an unknown reason, Pillow's Image.save function in mmds_to_imageai.py occasionally spits out unreadable data -- data that can't be parsed by Image.open. For that reason, this script finds the invalid files and deletes them. Bye-bye!

train_model.py

Trains a simple prediction model based on ImageAI's Prediction class

test_model.py

Tests the model against real-world images

directory structure

face-mask-detection-dataset/
    annotations/
    images/
medical-masks-dataset/
    medical-masks-dataset/
        labels/
        images/
cropped/
    mask/
    mask_weared_incorrect/
    none/
    poor/
    with_mask/
    without_mask/
aggregate/
    mask/
    improper/
    none/
training_data/
    json/
        model_class.json
    logs/
    models/
    test/
        mask/
        none/
        improper/
    train/
        mask/
        none/
        improper/

pip dependencies

Note: you must use Python <3.8 (I recommend 3.7).

tensorflow<2
scipy<1.5
numpy
keras
opencv-python
pillow
imageai

About

Fullstack application to tell if you're wearing a mask right

Resources

Stars

Watchers

Forks

Packages

No packages published

Languages