Deep face detection

Finding faces from DNN and Haar face detection.

Introduction

This work was to experiment face detection using OpenCV for the video anonymization.

In this repository, the all included methods were exploited from OpenCV examples including

1. Haar detector - Classic Haar cascade detector
2. DNN with no crop
3. DNN with crop
4. DNN with split videos by the ratio
   • Crop video by 1:1 ratio and slides the crop area to the rest
5. DNN with split videos by crop size
   • Split videos by less or more than 5x5 with overlapping area and apply method 4.

I suggest you to take a look at Video anonymization by FaceDetection-DSFD for the best performance and Video anonymization by lightDSFD for faster detection but less accurate.

Usage

usage: noone_video.py [-h] [--file FILE] [--out OUT] [--proto_file PROTO_FILE]
                      [--model_file MODEL_FILE] [--haar_file HAAR_FILE]
                      [--vertical VERTICAL] [--show SHOW]
                      [--debug_rect DEBUG_RECT]
                      [--debug_fill_color DEBUG_FILL_COLOR]
                      [--debug_text DEBUG_TEXT]
                      [--detector_dnn_nocrop DETECTOR_DNN_NOCROP]
                      [--detector_dnn_crop DETECTOR_DNN_CROP]
                      [--detector_dnn_scan DETECTOR_DNN_SCAN]
                      [--detector_dnn_crop_scan DETECTOR_DNN_CROP_SCAN]
                      [--detector_haar DETECTOR_HAAR]
                      [--detector_combined DETECTOR_COMBINED]
                      [--blur_faces BLUR_FACES] [--reduce_scale REDUCE_SCALE]
                      [--verbose VERBOSE] [--blur_level BLUR_LEVEL]
                      [cam_number]

Blurring faces on videos

Detailed arguments!

positional arguments:
  cam_number            Index number of the camera

optional arguments:
  -h, --help            show this help message and exit
  --file FILE           Video file path
  --out OUT             Output video path
  --proto_file PROTO_FILE
                        DNN proto file path
  --model_file MODEL_FILE
                        DNN model file path
  --haar_file HAAR_FILE
                        Haar cascade xml file path
  --vertical VERTICAL   0 : horizontal video(default), 1 : vertical video
  --show SHOW           0 : Hide figure, 1 : show figure
  --debug_rect DEBUG_RECT
                        Show rectangular around the detected faces. Default :
                        0
  --debug_fill_color DEBUG_FILL_COLOR
                        Fill colors on the candidate area of the faces,
                        Default : 0
  --debug_text DEBUG_TEXT
                        Show text information of the detectors, Default : 0
  --detector_dnn_nocrop DETECTOR_DNN_NOCROP
                        Use DNN without crop. Default : 0
  --detector_dnn_crop DETECTOR_DNN_CROP
                        Use DNN with crop. Default : 0
  --detector_dnn_scan DETECTOR_DNN_SCAN
                        Use DNN while scanning the frame. Default : 0
  --detector_dnn_crop_scan DETECTOR_DNN_CROP_SCAN
                        Use DNN while scanning the cropped frame. Default : 1
  --detector_haar DETECTOR_HAAR
                        Use Haar detector. Default : 0
  --detector_combined DETECTOR_COMBINED
                        Use combined detection result. Default : 0
  --blur_faces BLUR_FACES
                        Blur detected faces
  --reduce_scale REDUCE_SCALE
                        Reduce scale ratio. ex) 2 = half size of the input.
                        Default : 2
  --verbose VERBOSE     Show current progress and remaining time
  --blur_level BLUR_LEVEL
                        Blurriness of the detected face