Image classification method based on cross-type migration active learning

Abstract

The invention discloses an image classification method based on cross-type migration active learning. The method comprises the following steps of carrying out vectoring characteristic representation on images in auxiliary type data and target type data; constructing auxiliary type attribute representation and target type attribute representation; constructing a target function; optimizing and solving the target function to acquire a generation function; using the generation function and the target type attribute representation to acquire a classification model; using the classification model to calculate uncertainties of all the images without marks in the target type data; selecting the image without mark whose uncertainty is maximal and carrying out marking; updating a weight of an image in the target type data and reconstructing the target function. In the invention, an image sample who has the largest amount of information can be effectively selected in the target type data and is marked, which is good for training an accurate classification model under the condition that only a few of target types have marked data, and marking cost is reduced.

Description

technical field [0001] The invention relates to the field of image classification, in particular to an image classification method based on cross-category transfer active learning. Background technique [0002] With the large-scale growth of image data on the Internet, image classification technology has been widely concerned and applied. The existing image classification technology mainly trains the classification model of the target category through the method of supervised learning, that is, it is necessary to collect sufficient and high-quality labeled data for each classification model for model training. This method is suitable for simple classification tasks and scenarios with fewer categories. However, with the complexity of classification tasks, such as a large number of categories, category specialization, specialization, etc., the cost of collecting sufficient labeled data for the target category has greatly increased. Therefore, how to save labeling costs as mu...

AI Technical Summary

Problems solved by technology

[0004] Judging from the current research, the existing migration active learning methods can only deal with the situation where the categories of the target domain and the auxiliary domain are exactly the same. There is no solution for the two completely different, but the latter is more practical in practical applications common situation

Images