Face detector is a utility for locating faces in photographs. It can:
- Crop out the detected faces
- Provide a JSON file with the coordinates of the detected faces
- List the files that have faces
- Highlight the detected faces with bounding boxes
Face detector uses a combination of two algorithms to improve face detection performance. Initially it tries a deep neural network approach. A trained model called YuNet (MIT License) is included. This model generates few false positives but is less suited to large images with faces greater in size than 300x300 pixels. To solve this the image is progressively rescaled and face identification is repeated at each scale until a face is detected. If this approach fails a traditional Haar cascade algorithm is used. This is prone to generating false positives but is more successful with larger faces. To reduce the false positive rate, this approach is blended with the deep neural network to double check the accuracy.
Face detector can be useful for generating AI training data sets. It can produce faces with a square aspect ratio and at a specific size which training may require. Such data sets also often require duplicate images to be removed to avoid over-training, for which my photo de-duplication repo may helpful. It would be recommended to run photodedupe prior to detecting faces.
Face detector is written in Python. Simply install the requirements and then run the "face_detect.py" python script.
pip install -r requirements.txt
To crop out faces found in all photos contained with a directory:
./face_detect.py input_photos_dir/ --output-dir output_cropped_faces_dir/
To instead report the locations of faces in JSON format:
./face_detect.py input_photos_dir/ --json
Each of the detected faces will be quite closely cropped. To include more context around the face use the --expand_percent option. This accepts an integer which is a percentage expansion around the detected face e.g. this command line will double each width/height dimension (where possible) whilst keeping the face centred. If the face is at the edge of the image, it may not be possible to double the dimensions and the option will produce the largest expansion around the face that it can:
./face_detect.py input_photos_dir/ --output-dir output_cropped_faces_dir/ --expand-percent 100
AI training can require faces with a precisely square aspect ratio and a specific size. By default face detector can crop out rectangular regions. To force square images only use --square-aspect. AI training also often requires images of a specific size. To resize cropped images use the --resize option specifying a widthxheight. For example to produce images with a square aspect ration and an exact size of 256x256 pixels specify:
./face_detect.py input_photos_dir/ --output-dir output_cropped_faces_dir/ --square-aspect --resize 256x256
Instead of cropping faces, bounding boxes can also be drawn on the original image using the --highlight option e.g:
./face_detect.py input_photos_dir/ --output-dir output_bounding_boxes_dir/ --highlight
Faces below a certain size can also be ignored using --min-size option, which specifies a width/height below which faces will not be detected.
./face_detect.py input_photos_dir/ --output-dir output_cropped_faces_dir/ --min-size 100x100
Files that contain faces can also be listed on stdout. This is useful for piping a list of images that contain at least one face to another command e.g. to move all of the images that contain at least one face to another directory:
./face_detect.py input_photos_dir/ --file-list | xargs -i mv "{}" directory_of_images_with_faces/