Sequential (cascade) feature space reduction method to convert current space to linear space method is shown. Method is able to separate one point from other points using linear discriminant. Using the properties of separation in large spaces, every distribution of points could be convert to linear space statement is shown. It is shown that using these properties and the cascade reduction method it is possible to solve the face recognition problem in constant time, without cluster searching. Classification quality comparison of cascade reduction method with different face recognition methods is made. Experiment procedure and details of implementation is made, in particular types of convolutional neural networks that were used. The results shown by experiments says that is could be useful to use cascade reduction method to solve tasks with strong inference time requirements.

Authors: A. M. Golubkov, D. M. Klionskiy

Direction: Informatics, Computer Technologies And Control

Keywords: Cascade reduction, Principal Component Analysis, Linear Discriminant Analysis, face recognition, images classification

View full article