Dual-sampling integration classification model based on Fisher kernel

Abstract

The present invention provides a dual-sampling integration classification model based on Fisher kernel. Data is mapped to a high-dimensional Fisher space to obtain sample expression with better distinguishing features. Samples are subjected to multiple feature-level sampling in a Fisher space to obtain multiple view angles to increase information required in the classification so as to improve thestability of a base classifier. An integration method is employed to perform integration of inside of the view angles and outside of the view angles. Integration of different view angle expressions is employed to maintain the diversity of the whole sample. The whole system is designed, and a frame integration method is provided for an unbalance classification problem; the classification information is increased for the unbalance data through the system so as to provide a more accurate classification result; combined training models with different structures can be used for concrete problems according to the concrete problems; and different sample expression matrixes with various forms and vectors are generated according to the sample information quantity to enrich the training data so asto improve the final classification effect.

Description

technical field [0001] The invention relates to the technical field of pattern recognition, in particular to a Fisher kernel-based double sampling integrated classification model for unbalanced data classification. Background technique [0002] There are two main categories of methods in pattern recognition, generative models and discriminative models. The generation model focuses on the generation process of the probability density function, which can provide more information for the study of the data. Whereas discriminative models focus on direct classification and directly learn classification boundaries. It makes decisions based on the prior knowledge of the training set, so it is less computationally expensive than the generative model, and its performance is better than the generative model. [0003] Existing methods to solve the imbalance problem mainly include sampling method, method of introducing cost function, and integration method. The sampling method rebalan...

AI Technical Summary

Problems solved by technology

Usually in imbalanced problems, the difference in sample size leads to the fact that minority samples are rarely sufficiently minimized

Images