Deep Convolutional Neural Networks and Maximum-Likelihood Principle in Approximate Nearest Neighbor Search

Deep convolutional neural networks are widely used to extract high-dimensional features in various image recognition tasks. If the count of classes is relatively large, performance of the classifier for such features can be insufficient to be implemented in real-time applications, e.g., in video-based recognition. In this paper we propose the novel approximate nearest neighbor algorithm, which sequentially chooses the next instance from the database, which corresponds to the maximal likelihood (joint density) of distances to previously checked instances. The Gaussian approximation of the distribution of dissimilarity measure is used to estimate this likelihood. Experimental study results in face identification with LFW and YTF datasets are presented. It is shown that the proposed algorithm is much faster than an exhaustive search and several known approximate nearest neighbor methods.