Deep neural networks and maximum likelihood search for approximate nearest neighbor in video-based image recognition

We analyzed the way to increase computational efficiency of video-based image recognition methods with matching of high dimensional feature vectors extracted by deep convolutional neural networks. We proposed an algorithm for approximate nearest neighbor search. At the first step, for a given video frame the algorithm verifies a reference image obtained when recognizing the previous frame. After that the frame is compared with a few number of reference images. Each next examined reference image is chosen so that to maximize conditional probability density of distances to the reference instances tested at previous steps. To decrease the required memory space we beforehand calculate only distances from all the images to small number of instances (pivots). When experimenting with either face photos from Labeled Faces in the Wild and PubFig83 datasets or with video data from YouTube Faces we showed that our algorithm allows accelerating the recognition procedure by 1.4–4 times comparing with known approximate nearest neighbor methods.