RHF: ResNet50-HDC-FPN-FasterRCNN-CMFD-LCMFD

简体中文 | English

Paper

Wang, S., Wang, X., & Guo, X. (2023). Advanced Face Mask Detection Model Using Hybrid Dilation Convolution Based Method. Journal of Software Engineering and Applications, 16(1), 1-19.

Use code, data, weights, etc... Please cite 💝

@article{wang2023advanced,
  title={Advanced Face Mask Detection Model Using Hybrid Dilation Convolution Based Method},
  author={Wang, Shaohan and Wang, Xiangyu and Guo, Xin},
  journal={Journal of Software Engineering and Applications},
  volume={16},
  number={1},
  pages={1--19},
  year={2023},
  publisher={Scientific Research Publishing}
}

Model weighting file

• Baidu: https://pan.baidu.com/s/1oW3BopexHkJdsQlb79hsdw Extraction code: 2swh
• Dropbox: https://www.dropbox.com/s/rg7dqkr71bylaey/save_weights.zip?dl=0
• Google Drive: https://drive.google.com/file/d/1-v7t9nGHauiUbF_o69d0gz52LSry1vSA/view?usp=share_link

Dataset files CMFD & LCMFD

• Baidu: https://pan.baidu.com/s/1TLEdIfqfQXI-PT49Snv3aQ Extraction code: 1111

Environment configuration.

• Python3.6/3.7/3.8
• Pytorch1.7.1
• pycocotools(Linux:pip install pycocotools; Windows:pip install pycocotools-windows)
• Partial code reference: https://github.com/WZMIAOMIAO/deep-learning-for-image-processing

File structure.

• backbone: feature extraction network ResNet50
• backbone_hdc: ResNet50-HDC The hybrid inflated convolution rate integrated residual network of this paper
• network_files: Faster R-CNN network (including Fast R-CNN and modules such as RPN)
• train_utils: training validation related modules (including cocotools)
• my_dataset.py: custom dataset for reading VOC datasets
• train_mobilenet.py: use MobileNetV2 as the backbone for training
• train_resnet50_fpn.py: use resnet50 + FPN as backbone for training
• train_resnet50_hdc_fpn: train with resnet50-HDC + FPN as backbone
• train_multi_GPU.py: train with multiple GPUs
• predict.py: Simple prediction script to perform prediction tests using trained weights
• validation.py: validate/test the COCO metrics of the data using the trained weights, and generate record_mAP.txt file
• classes.json: pascal_voc tag file

Pre-trained weights download address (download and put in backbone folder).

• MobileNetV2 backbone: https://download.pytorch.org/models/mobilenet_v2-b0353104.pth
• ResNet50+FPN backbone: https://download.pytorch.org/models/fasterrcnn_resnet50_fpn_coco-258fb6c6.pth
• Note that the downloaded pre-training weights should be renamed, e.g., the fasterrcnn_resnet50_fpn_coco.pth file is read in train_resnet50_fpn.py Not fasterrcnn_resnet50_fpn_coco-258fb6c6.pth

dataset

• LCMFD

  • num of all_annotations: 8232
  • num of with_mask: 3229
  • num of poor_mask: 2813
  • num of none_mask: 2190

• CMFD:MaskDatasets_Augment

  • num of all_annotations: 131422
  • num of with_mask: 53039
  • num of poor_mask: 47203
  • num of none_mask: 31180

Training method

• Note that the dataset file name in *.py is changed to the dataset file name on your computer
• Modify the parameters in train_res50_fpn.py train_res50_hdc_fpn.py
  • Modify train_res50_fpn.py line 185 to replace data_path
  • Modify train_res50_hdc_fpn.py line 185 Replace data_path
• Make sure the backbone folder contains pre-trained model weights
• If you want to train mobilenetv2+fasterrcnn, use the train_mobilenet.py training script directly
• To train resnet50+fpn+fasterrcnn, use the train_resnet50_fpn.py training script directly
• To train resnet50-hdc+fpn+fasterrcnn, use the train_resnet50_hdc_fpn.py training script directly
• To train with multiple GPUs, use the python -m torch.distributed.launch --nproc_per_node=8 --use_env train_multi_GPU.py command, with the nproc_per_node parameter being the number of GPUs used
• If you want to specify which GPU devices to use you can prefix the command with CUDA_VISIBLE_DEVICES=0,3 (e.g. I just want to use the 1st and 4th GPU devices in the device)
• ```CUDA_VISIBLE_DEVICES=0,3 python -m torch.distributed.launch --nproc_per_node=2 --use_env train_multi_GPU.py``

Caution

• When using the training script, be careful to set '--data-path' (mask_root) to the root directory where you store the 'MaskDatasets_Augment' or 'MaskDatasets_NotAugment' folder
• Since Faster RCNN with FPN structure is very memory hungry, if the GPU memory is not enough (if the batch_size is less than 8), it is recommended to use the default norm_layer in the create_model function. If the GPU memory is not enough (if the batch_size is less than 8), it is recommended to use the default norm_layer in the create_model function.
• When using the prediction script, set 'train_weights' to your own generated weight path.
• When using the validation file, take care to make sure that your validation set or test set must contain targets for each class, and use it with only '--num-classes', '--data-path' and '--weights', and leave the rest of the code unchanged as much as possible