当前位置：网站首页>Mask wearing face data set and mask wearing face generation method

Mask wearing face data set and mask wearing face generation method

2022-06-29 00:39:00 【pan_ jinquan】

Mask wearing face data set and mask wearing face generation method

1. Mask wearing face data set

（1） Open source datasets - Mask wearing face data set

（2） Generating data sets of face masks

2. Face generation method with mask

（1） The idea of generating face masks

（2） Mask template

（3） Face generation with mask Demo

（4） Generate face effect display with mask

3. Generate a mask data set and download the source code

4. Mask wearing face detection and mask wearing recognition ( contain Python Android Source code )

At present, many free face recognition programs have been opened on the Internet / Face detection dataset , For example, the common LFW,FDDB,WIDER FACE,300W etc. , I guess it adds up to several memories . But most of these face data are faces without masks , It can not be directly used in face mask recognition . There are also a few open source face data sets with masks on the Internet , I spent some time , Sort out the currently common face data sets with masks , And the synthesis of face data with masks / Generation method .

contain 5 Data sets ： facemask-train1, facemask-train2,facemask-train3, synthetic-train1,synthetic-train2 ,facemask-test , There are about 50000+ The data of ：
Generate mask wearing face code ： python create_facemask.py

Project data and generate the source code of face masks ： Mask wearing face data set and generating mask wearing face data

1. Mask wearing face data set

（1） Open source datasets - Mask wearing face data set

Some open source , Free face mask data set ,download Come down , It is found that these open source data are dirty , A lot of them are marked incorrectly , I need to clean myself ~

Data sets	explain
virus-mask-dataset	There are three types of labels in the dataset :mask This person is wearing a virus resistant mask ;nomask No mask ;wrongmask This person is wearing a mask that does not prevent viruses ; The marking method is : Mark the whole head of the human body + Upper shoulder Address ：https://github.com/hikariming/virus-mask-dataset
MaskedFace-Net	Correctly occlude the face data set (CMFD), Error occlusion face data set (IMFD) And global occlusion face detection (MaskedFace-Net) The combination of . Address ：https://github.com/cabani/MaskedFace-Net
Real-World Masked Face Datase	Address ：https://github.com/X-zhangyang/Real-World-Masked-Face-Dataset explain ：https://blog.csdn.net/HyperAI/article/details/105154132

（2） Generating data sets of face masks

Most of the face data on the Internet are faces without masks , It can not be directly used in face mask recognition . In view of this , We can consider our own synthesis / Generate face data with masks , The following is my collection and collation of face data sets with masks and synthetic data sets , There are about 50000+ The data of ：

Data sets	explain
facemask-train1	Face data set with masks collected from the Internet ( Such as virus-mask-dataset), about 7000+ A picture , And cleaned some samples with wrong marks Every picture is marked mask( Wear a mask ) and nomask( No mask ) The detection frame of The dimension format is standard VOC xml Format , Can be used for Face detection training data use The face area has been cropped , And cleaned some samples with wrong marks ; among mask( Wear a mask ) Face has 3000+ Zhang ,nomask( No mask ) Face has 10000+ Zhang , It can be used as classification training data set ,
facemask-train2	Face data set with masks collected from the Internet , about 3500+ A picture , Every picture is marked mask( Wear a mask ) and nomask( No mask ) The detection frame of The dimension format is standard VOC Format , However, the face frame marked is relatively large , It is not recommended to use face detection training data The face region image has been cropped , And cleaned some samples with wrong marks ; among mask( Wear a mask ) Face has 2000+ Zhang ,nomask( No mask ) Face has 6000+ Zhang , It can be used as classification training data set
facemask-train3	Face data set with masks collected from the Internet , among mask( Wear a mask ) Face has 600+ Zhang ,nomask( No mask ) Face has 1700+ Zhang , It can be used as classification training data set The original images have been cropped into face images , therefore Not suitable for face detection ; It can be used as classification training data set
synthetic-train1	This is the synthetic face data with mask among mask( Wear a mask ) Face has 7000+ Zhang ,nomask( No mask ) Face has 7000+ Zhang , It can be used as classification training data set
synthetic-train2	This is the synthetic face data with mask among mask( Wear a mask ) Face has 6000+ Zhang ,nomask( No mask ) Face has 6000+ Zhang , It can be used as classification training data set
facemask-test	This is the face test set with mask among mask( Wear a mask ) Face has 300+ Zhang ,nomask( No mask ) Face has 300+ Zhang , Used for classification model testing

2. Face generation method with mask

（1） The idea of generating face masks

First, face detection and face key point detection
According to the face key , Calculate the width of the cheek and the position of the bridge of the nose , Empathy , It is necessary to calculate the size of the outlet cover template 、 Direction and position ;
Finally, paste the mask template on the face , Wait until the synthetic mask face image

Some masks generate code references ：

GitHub - sevenHsu/FaceMask_CelebA: Simulated masked face with dataset CelebA

Put a mask on your face ,Python The actual combat project is coming _ I love Python Data mining blog -CSDN Blog

I tested the current online generated code for wearing masks , For faces that have been corrected , The generation effect is good , But if the original face itself is tilted ( Crooked head ), In this case, use the original code directly , The effect of synthesis is very poor , On this basis , I did optimization work ：

Added mask template , share 29 Different types of mask templates , Provide diversity of synthetic mask data
The inclination angle of the mask is optimized , Make the face tilt ( Crooked head ) when , The synthetic effect is also good
Optimized face detection , Using lightweight face detection , Speed up generation

Original picture	The original code generates wearing masks	After optimization, mask wearing will be generated

For faces that have been corrected , The generation effect is good , But if the original face itself is tilted ( Crooked head ), In this case, use the original code directly , The effect of synthesis is very poor

（2） Mask template

Mask template is the mask Mask Images （PNG Format ）, Can pass Photoshop A stingy way , Remove the background ; A total of 29 Different types of mask templates , Ensure the diversity of mask data .

（3） Face generation with mask Demo

Install well python Dependent package

pybaseutils
opencv-python==4.5.1.48
face_recognition
tqdm
torch
torchvision

The code already comes with test images and data , You can also change to your own face data set ：

python create_facemask.py

# -*-coding: utf-8 -*-
"""
    @Author : panjq
    @E-mail : [email protected]
    @Date   : 2022-06-25 22:23:11
    @Brief  :  Generate face data set with mask 
"""
import os
import sys

sys.path.insert(0, os.getcwd())
sys.path.insert(0, "libs")
from tqdm import tqdm
from facemask.wearmask import FaceMaskCreator
from pybaseutils import file_utils, image_utils


class FaceMaskDemo(object):
    def __init__(self):
        self.mask_creator = FaceMaskCreator(detect_face=True, alignment=False)

    def create_wear_mask_faces(self, image_dir, out_dir=None, vis=True):
        """
         Generate face data set with mask 
        :param image_dir:  Face image directory 
        :param out_dir:   Generate mask face output directory 
        :param vis:  Whether the visual effect 
        :return:
        """
        image_list = file_utils.get_files_lists(image_dir)
        for image_path in tqdm(image_list):
            image_id = os.path.basename(image_path).split(".")[0]
            image = image_utils.read_image(image_path, size=(512, None), use_rgb=True)
            mask, face_rects = self.mask_creator.create_masks(image, mask_type="random", vis=vis)
            if out_dir:
                self.mask_creator.save_image(image, mask, face_rects, out_dir, image_id)


if __name__ == '__main__':
    image_dir = "./facemask/test_image"  #  Face pictures 
    out_dir = "./output"  #  Generate mask face output directory 
    fm = FaceMaskDemo()
    fm.create_wear_mask_faces(image_dir, out_dir, vis=True)

（4） Generate face effect display with mask

Original picture	Figure of synthetic belt mask

3. Generate a mask data set and download the source code

The download address contains ：

https://mp.weixin.qq.com/s/4dtC9CeP50M-3nn3xel1Ow

contain 5 Data sets ： facemask-train1, facemask-train2,facemask-train3, synthetic-train1,synthetic-train2 ,facemask-test , There are about 50000+ The data of ：
Generate mask wearing face code ： python create_facemask.py