Incremental picture classification method based on semi-supervised learning

Abstract

The invention provides an incremental picture classification method based on semi-supervised learning, and the method comprises the steps: 1, carrying out the preprocessing of a new type of pictures and a public data set picture (serving as auxiliary data) appearing in streaming data, and respectively putting the pictures into a set A and a set B; 2, generating a copy for the current model, and fixing copy model parameters; 3, training a teacher model by using the new data; 4, in the current model, for each new category, adding a header network to form a new model of the to-be-updated parameters; 5, performing shuffle operation on the auxiliary data and the new data to integrally form final training data; 6, calculating a target vector for the training data by using the model copy and theteacher model; 7, inputting the training data into the new model, and calculating a loss value according to a model output result and the target vector. 8, adjusting model parameters by using a gradient descent algorithm; and 9, testing the prediction precision of the model.

Description

technical field [0001] The invention relates to an incremental picture classification method based on semi-supervised learning. Background technique [0002] With the development of deep learning, major breakthroughs have been made in many related research fields and practical applications. One of the research directions is image classification. However, the training of the current deep learning model must require all the training data to be sent to the model for training in advance, and incremental learning cannot be performed. However, in practical applications, it is impossible to obtain complete data at one time, especially for streaming data scenarios. This requires the deep learning model to be able to continuously learn and update its own parameters based on the new data provided later. In the learning process of incremental classification, the model needs to be able to learn new knowledge while avoiding the forgetting of old knowledge—this is the catastrophic forget...

AI Technical Summary

Problems solved by technology

[0006] Purpose of the invention: In order to solve the problem of forgetting disaster in the process of class incremental learning, while having both space efficiency and time efficiency, this invention proposes an incremental image data classification method based on semi-supervised learning, and improves the accuracy of the model by using auxiliary data. The memory ability of old knowledge, thereby improving the overall classification accuracy of the model

Images