New face detection algorithm can spot even partially hidden faces
Yahoo! Labs and Stanford University are developing a fast and simple facial recognition system as part of a new research project, according to The Register.
The researchers have developed what they call the Deep Dense Face Detector, which is designed for “minimal complexity” and is able to identify faces that are partially occluded or rotated.
Sachin Sudhakar Farfade and Mohammad Saberian of Yahoo! and their colleague Li-Jia Li (Yahoo! and Stanford) have developed a new technique that takes advantage of “deep convolutional neural networks”, or a network in which individual neurons are spread out in a manner that they respond to overlapping regions in the visual field.
The researchers took annotated images of faces from various angles to serve as the algorithm’s training set.
The set’s 21,000 images with 24,000 annotations, along with random flipping of images, resulted in 200,000 “positive” examples (images with faces) and 20 million “negative examples (images with no faces).
The AlexNet model was applied to 50,000 iterations across groups of 128 images that included 32 positive and 96 negative examples.
Since the method does not require annotation based on poses or landmarks, the researchers said that it is possible to “detect faces in all orientations using a single model”.
According to a report by Technology Review, if the algorithm works, it will allow a large repository of past photos, CCTV and video footage completely searchable.