Project: Mask Image Classification

Click for Detail process in wrap-up report

The problem of classifying a given person image by 3-criteria

• Mask Wear Form: Not wear, Incorrect, Wear
• Gender: Male, Female
• Age: 0 to 29, 30 to 59, 60~

Model Structure

(1) Without making 3 separate models to determine person’s mask wear state, age, and gender, suggest a new end-to-end model by creating three heads on CNN backbone for each prediction, which makes better performance.

(2) Fine Tuning with 3 additional Heads(each head for prediction of mask wear form, gender, and age) with 0.1 * learning rate and Pre-trained CNN Backbone Models with 1.0 * learning rate

• Pre-trained Backbone Model: ResNet50, EfficientNet-b4, ResNest200e, Vit

Task Description

With the spread of COVID-19, people in our country as well as around the world have many restrictions on economic and productive activities. The most important thing to prevent the spread of infection is to cover everyone's nose and mouth with a mask to block the transmission route from any possible infected person.

To do this, people in public places must wear masks, and above all, it is important to wear them properly so that they can cover their nose and mouth completely. However, additional human resources will be needed to check everyone's correct wearing of masks in large public areas.

Therefore, we need a system that can automatically identify whether this person is wearing a mask, not wearing it, or wearing it correctly, just by the image of the person's face shown on the camera. If this system is in place at the entrance to public spaces, it will be sufficiently inspected with a small amount of human resources.

Requirements

Confirmed that it runs on Ubuntu 18.04, Python 3.7, and pytorch 1.7.1.

The necessary packages can be installed with the command below.

pip install -r requirements.txt

Getting Started

Step 1: Download Source Code

• augmentation.py: custom augmentation methods for training
• loss.py: various loss functions for backward propagation (classification)
• optimizer.py: various optimizers for updating parameters
• scheduler.py: various schedulers for learning rate
• pytorch_tools.py: using early stopping
• preprocess.py: make new data csv file for correct validation
• mask_dataset.py: pytorch dataset for both training & Inference
• mask_model.py: CNN models using various CNN backbone
• train.py: training model with using various hyperparmeters
• inference.py: make prediction by trained models

Step 2: Preparing Datset

+-- train/
|   +-- images/
|       +-- 000001_female_Asian_45/
|       +-- 000002_female_Asian_52/
|       +-- …
|   +-- train.csv
+-- public/
|   +-- images/
|       +-- 814bff668ae5b9c595ceabcbb6e1ea84634afbd6.jpg
|       +-- 819f47db0617b3ea9725ef1f6f58e56561e7cb4b.jpg
|       +-- …
|   +-- info.csv
+-- private/
   +-- images/
       +-- 0001b62fb1057a1182db2c839de232dbef0d1e90.jpg
       +-- 10006ee2ebd5a23196ee3bda4d7e06489ed40ce7.jpg
       +-- …
   +-- info.csv

preprocessing.py makes new csv file for preventing data leakage when splitting train-validation dataset

python3 ./preprocess.py

Step 3: Training Model with various hyperparameters

python3 ./train.py
--csv_path="train_st_df_58.csv"
--lr=0.006
--batch_size=64
--nworkers=4
--nepochs=70
--seed=42
--patience=30
--resize_width=380
--resize_height=380
--model="EfficientNetMaskClassifier"
--kfold=2
--train_augmentation="ResizeVariousAugmentation" 
--test_augmentation="BaseAugmentation"
--optimizer="SGDP"
--criterion="label_smoothing"
--scheduler="OneCycleLR"
--description="age filter + SGDP + Data Aug + label smoothing(for mask & age) + onecycleLR"

Step 4: Make Predictions

python3 ./inference.py
--batch_size=1
--nworkers=4
--seed=42
--model="EfficientNetMaskClassifier"
--test_augmentation="BaseAugmentation"
--model_version=0408_1400
--model_epoch=27