The goal of this repo is to make a tool to quickly and easily de-identify a large number of images that is free to use even for commercial use.
This takes a directory of images and removes the faces from the images. It uses the Tensorflow Object detection API and used a pretrained faster_rcnn_resnet101 model that was retrained using hand-labeled images that were marked for commercial use.
In the future, I would also like it to block out or remove images from the output that have tattoos but I was not successful in my initial attempt.
On my machine, with a Titan X(Pascal) on a SATA SSD it takes about 5.8 seconds for tensorflow to get up and running and about 0.28 seconds per image. This will probably vary with hardware and image size.
To use this tool install the Tensorflow Object detection API, clone the repo and run the image_detect.py followed by the relative path to the directory for the input images and a relative path for the output images that have the faces removed. The defaults are input/ and output/
python3 image_detect.py --input_images_path=input --output_images_path=output/
Here are some examples of it in action!
