Real-time Medical Mask Detection Based on Faster RCNN

This is an official implementation of blog: website-under-construction

Dataset

We use a small dataset Mask Detection, found available on Kaggle. In this dataset, we have 853 images belonging to 3 classes ('with mask', 'without mask' and 'masks worn incorrectly'). And the whole dataset is stored in data_augmentation/input.

Data Augmentation

With files(images and annotations) to augment put in input/(in two sub folders). Create a output folder under directory of '../data_augmentation' with two sub-folders images and annotations in it. The relative locations of files are as follows：

|--data_augmentation
   |--util
   |--augment.py
   |--input
   |--output
      |--images
      |--annotations      

Change the INPUT_DIR and OUTPUT_DIR in augment.py to corresponding paths. AUGMENT_SIZE decides how many augmented images will be produced from one original image.

INPUT_DIR = '../data_augmentation/input'
OUTPUT_DIR = '../data_augmentation/output'
AUGMENT_SIZE = 4

Run script:

python ../data_augmentation/augment.py

Augmented images and annotations are generated in output/.

Requirements

Python 3.8 or later with all requirements.txt dependencies installed, including torch>=1.7. To install run:

pip install -r requirements.txt

Train

Open mask-fasterrcnn.ipynb with Colab. Change data_path and output_path in augment.py to corresponding paths, and run blocks in the notebook.

data_path = r'../data_augmentation/output'
output_path = r'../results'

Real-time Mask Detection

Since colab cannot connect to the local camera on your computer, you have to build a local environment. After that, change model_path to corresponding local path:

model_path = r'C:\Downloads\Epoch_24_model.pt'

Run script:

python C:\Downloads\Medical-Mask-Detection-Based-on-Faster-RCNN\realtime\camera.py

However, the processing speed is much slower than our expectation, which only achieves to detect one frame for each five seconds. The way for feeding frame into the model have a lot of room to be optimized. For more details, please check the source code.

About me

EDUCATION:

• Bachelor of Communication Engineering (2022.7 Expected)
• Qiushi Honor College, Tianjin University (TJU)

CONTACT ME:

• E-mail: [email protected]
• Address: 92 Weijin Road, Tianjin, P.R. China