Face detection and blurring methods and systems
Methods, systems, and devices for generating a training set and using the training set to train an artificial intelligence algorithm to detect and blur faces in images of an interior space are described. An example method for generating a training set includes selecting a first image depicting an interior of a room, selecting a second image depicting at least a face of a person, generating, based on a transparency parameter and a roughness parameter, a semi-transparent surface image, selecting an illumination parameter to configure a first illumination level for the first image and a second illumination level for the second image, and generating a first composite image of the training data set by combining the illumination parameter, the first image, the second image, and the semi-transparent surface image.
1. A method of face detection and blurring, comprising:
receiving a plurality of images that captures a person traversing an interior of a room;
processing, using a face detection algorithm, the plurality of images to identify an anchor image comprising a complete view of a face of the person and generating a bounding box for the face of the person;
processing, using an optical region tracking algorithm, the plurality of images to identify a trajectory of the bounding box across the plurality of images;
applying a blurring kernel over the bounding box across each of the plurality of images;
determining a subset of images of the plurality of images, wherein a position of the bounding box is on the trajectory in each of the subset of images; and
computing, based on the subset of images, a probability corresponding to the bounding box following the trajectory, wherein applying the blurring kernel is based on a comparison of the probability to a threshold,
wherein the trajectory comprises a backward-in-time trajectory that tracks the bounding box from the anchor image to a first image of the plurality of images and a forward-in-time trajectory that tracks the bounding box from the anchor image to a last image of the plurality of images,
wherein the optical region tracking algorithm refrains from using the face detection algorithm,
wherein the trajectory tracks the bounding box based on a pixel-wise, frame-to-frame similarity of a region associated with the bounding box, and
wherein at least the first image or the last image comprises a partial view of the face of the person and a partial bounding box.
2. The method of claim 1 , wherein the blurring kernel is a Gaussian blurring kernel.
3. The method of claim 1 , wherein the optical region tracking algorithm comprises a channel and spatial reliability tracking (CSRT) algorithm.
4. The method of claim 1 , wherein the face detection algorithm uses a neural network.
5. The method of claim 4 , wherein the first image is a first-in-frame image comprising the partial view of the face of the person, and wherein the last image is a last-in-frame image comprising the partial view of the face of the person.
6. The method of claim 1 , wherein a kernel size of the blurring kernel is equal to 0.5×W, and wherein W is a width of the bounding box.
7. The method of claim 1 , wherein the threshold is a user-configured predetermined threshold.