Face Recognition in Real-Time Applications: A Comparison of Directed Enumeration Method and K-d Trees

The problem of face recognition with large database in real-time applications is discovered. The enhancement of HoG (Histogram of Gradients) algorithm with features mutual alignment is proposed to achieve better accuracy. The novel modification of directed enumeration method (DEM) using the ideas of the Best Bin First (BBF) search algorithm is introduced as an alternative to the nearest neighbor rule to prevent the brute force. We present the results of an experimental study in a problem of face recognition with FERET and Essex datasets. We compare the performance of our DEM modification with conventional BBF k-d trees in their well-known efficient implementation from OpenCV library. It is shown that the proposed method is characterized by increased computing efficiency (2-12 times in comparison with BBF) even in the most difficult cases where many neighbors are located at very similar distances. It is demonstrated that BBF cannot be used with our recognition algorithm as the latter is based on non-symmetric measure of similarity. However, we experimentally prove that our recognition algorithm improves recognition accuracy in comparison with classical HoG implementation. Finally, we show that this algorithm could be implemented efficiently if it is combined with the DEM.