Small sample class incremental learning method based on representation prediction

Abstract

The invention discloses a small sample class incremental learning method based on representation prediction, and relates to the technical field of image recognition and classification, and the method comprises the following steps: S1, completing the basic task training of base class data Dbase through a feature extractor f and a softmax classifier structure, and obtaining the features of the base class data through the basic task training; S2, after the basic task training is completed, removing the softmax classifier, inputting the features output by the feature extractor f into an NCM classifier, and calculating the class center of each class; S3, obtaining the output of a sample in the new class DN through a feature extractor f, comparing the similarity between the obtained sample output and the class centers of all the base classes obtained in the step S2, predicting the features of the sample according to the similarity degree, and normalizing the output similarity value as the weight of the corresponding class center ub. Through the design, the purpose of classifying all the learned categories under the condition that the categories are continuously increased is achieved.

Description

technical field [0001] The invention relates to the technical field of machine learning, in particular to a small-sample incremental learning method based on representation prediction. Background technique [0002] With the development of information technology, data has increased rapidly, and the demand for data processing has also greatly increased or decreased. In machine learning, the identification and classification of data is mainly to extract specific features in the data, represent the information of the data through specific features, and then identify and classify the data according to the extracted specific features. Machine learning can be used in many fields, and machine learning can quickly and accurately extract the characteristics of objects under various complex conditions, which has a wide range of application prospects. In the existing machine learning classification methods, fixed data can be identified and classified , but when a small number of new sa...

AI Technical Summary

Problems solved by technology

Machine learning can be used in many fields, and machine learning can quickly and accurately extract the characteristics of objects under various complex conditions, which has a wide range of application prospects. In the existing machine learning classification methods, fixed data can be identified and classified , but when a new small number of samples appear, the system cannot recognize

Images