Method of outliers removal based on the weighted training samples of w-objects

Abstract

The problem of preprocessing training samples to improve the efficiency of trainable recognition systems is considered in the paper. A new method for solving the problem of outliers removal based on constructing weighted reduced samples of w-objects is proposed. The wGridDC method for constructing the weighted sample of w-objects by superimposing the grid features on the space and constructing weighted objects of new sample by analyzing the contents of cells is used as a basis for the proposed method.Within the proposed method, two outliers removal algorithms are developed. The algorithm for constructing the weighted training sample of w-objects with simultaneous outliers removal at a given filtering threshold is focused on the use in the tasks that require not only filtering the original data, but also controlling the size of the sample. Herewith, filtering threshold is user-defined. The algorithm for constructing the weighted training sample of w-objects with simultaneous outliers removal at automatic filtering threshold detection is focused on the tasks that require constructing samples, providing the highest efficiency of the system.Analysis of the effectiveness of the proposed method has shown that the main advantage of the threshold filtering algorithm is the ability to control the size of the sample. The main advantage of the non-threshold filtering algorithm is the ability to automatically select the value of the filtering threshold that provides the greatest efficiency of the recognition system as a whole. Thus, the proposed method in general and both its constituent algorithms allow to obtain the samples, providing high efficiency of trainable recognition systems.