Yahoo! Boffins! Improve! Face! Search!

A research collaboration between Yahoo! Labs and Stanford University has yielded a fast and simple approach to facial recognition.

If you're the kind of person that finds features like being tagged in images on Facebook unsettling, you might not see this as good news: not only do the researchers' Deep Dense Face Detector claim “minimal complexity”, it can identify faces that are partially occluded or rotated.

The technique used by Sachin Sudhakar Farfade and Mohammad Saberian of Yahoo! and their colleague Li-Jia Li (Yahoo! and Stanford), their approach was developed to take advantage of what's called “deep convolutional neural networks”.

They took annotated images of faces from many angles as the algorithm's training set: 21,000 images with 24,000 annotations, along with random flipping of images, yielded 200,000 “positive” examples (images with faces in them). To these were added 20 million images with no faces (negative).

The AlexNet model was run through 50,000 iterations across batches of 128 images containing 32 positive and 96 negative examples.

Because the method doesn't need annotation based on poses or landmarks, the researchers write that it's able to “detect faces in all orientations using a single model”.

The researchers reckon they can way outpace existing approaches to facial recognition

As Technology Review notes, if the algorithm works, it won't just be applicable to new photos: it'll make a huge history of photos, CCTV and video footage searchable.

The paper is at Arxiv here. ®