FaceMask-Detection system using YOLOv3 in Keras

First of all, I must mention that this code used in this project originally is not mine. I was inspired by qqwweee/keras-yolo3, he wrote this code to train custom YOLOv3 Keras model.

Few changes are:

1. To load images for detection, replaced PIL with opencv.
2. Used the same opencv font for drawing text on detected images.
3. Wrote seperate detection script for images and videos.

Dependencies :

• Python 3.6.10
• Keras 2.2.4 conda install keras=2.2.4
• Tensorflow 1.15 conda install tensorflow-gpu=1.15 or conda install tensorflow
• OpenCV-python
• PIL
• NumPy
• Moviepy(for the detection of videos) pip install moviepy
• LabelImg

Note that keras and tensorflow have their own dependencies. I recommend using Anaconda for handlinng the packages.

Dataset :

I created the dataset for this project. It is divided into two categories ie "with_face_mask" and "without_face_mask". I downloaded the images for both categories from internet and saved them in respective folder in the local disk.

The most important part is the creation of Annotations and bounding boxes; for that purpose I used the LabelImg annotation tool and created the labels and bounding boxes for each image.

In LabelImg, the annotations files are available in two format ie VOC format(XML file) or YOLO format. So, I got the data annotation file in YOLO format directly from labelImg ie dataset_train.txt file.

Usage :

• Either download YOLOv3 weights from YOLO website & convert them, or download the YOLOv3 weights in keras format from release section.
  Copy downloaded weights file to "model_data" folder.
• Run the "train.py" file for training the model. (I opt batch_size = 4, increase the batch_size if have more memory).
  I tried to train with GTX1650, received memory error while using "batch_size = 32 or 16", was able to train with 4.
  Increase epochs count for better model accuracy.
• After finishing training, in "trained_weights/003/" folder there should be created new "trained_weights_final.h5" model file.
  This will be used for custom detection.
• Test YOLOv3 with images, use "image_detect.py"
  Simply change "model_path" and "classes_path", to the used files. Then at the end of file, change the test image name and directory.
• Test YOLOv3 with videos, use "video_detect.py"
  Simply change "model_path" and "classes_path", to the used files. Then at the end of file, just change the video and video_output directory.

Screenshots :

Video

I've also checked the final model on a video file. The output file can be seen here.