A method and system for semi-supervised classification of multi-class images

Abstract

The invention discloses a multi-class image semi-supervised classification method and system. Firstly, similarity learning is performed on labeled image samples and unlabeled image samples in the training set, and a similar neighbor graph and normalized weights are constructed to represent sample similarity. , and then initialize a class label matrix. In order to effectively reduce the influence of mixed signals in the "soft category label" prediction label F on the results, l2,1‑norm regularization is introduced, and a non-negative and column sum of 1 is applied to F Constraints to ensure that the estimated "soft label" meets the probability definition and non-negativity, and finally use the parameters to weigh the impact of the similarity measure, the initial category label and the l2,1-norm regularization on the classification, complete the semi-supervised learning modeling, take The maximum value of similarity probability is used for image category identification to obtain classification results. By introducing l2,1-norm regularization, the influence of mixed signals on classification is reduced, and the classification accuracy is improved. In addition, it can also effectively classify data outside the training set, and has good scalability.

Description

technical field [0001] This application relates to the technical fields of data mining, machine learning and pattern classification, and in particular to a method and system for semi-supervised classification of multi-category images. Background technique [0002] With the advent of the era of information and data explosion, classification technology has become one of the most important research topics in the fields of data mining and pattern recognition. Classification mainly realizes the classification of unknown categories of data. It is of great significance in the fields of medical data analysis, credit grading of credit cards and image classification. Once the research is successful and put into application, it will produce huge social and economic benefits. However, most of the data in the real world (such as images in the Internet) have no class labels, and the manual calibration process of samples is very time-consuming, laborious and expensive, which makes the accu...

AI Technical Summary

Problems solved by technology

However, there are still some shortcomings in the existing methods, for example, how to effectively remove the mixed signals in the predicted "soft class label" prediction matrix and its impact on the classification results, and whether the output soft class label information always meets the definition of probability (that is, the probability sum is 1 and non-negative), etc.

Images