Semi-supervised image classification method based on joint training generative adversarial network

Abstract

The invention discloses a semi-supervised image classification method based on a joint training generative adversarial network. The method comprises the following steps:1, setting the generative adversarial network; 2, dividing a label data set L and a label-free data set U; 3, training a generator G; 4, training a discriminator D1 and a discriminator D2, and iteratively updating and expanding thelabel sub-sample set; 5, obtaining a trained generative adversarial network; and step 6, classifying the test set by using the trained generative adversarial network. According to the invention, thediscriminator D1 and the discriminator D2 are adopted for combined training, so that the influence of distribution errors of a single discriminator on the generative adversarial network is reduced; according to the joint training-based generative adversarial network, the dependence of the generative adversarial network on label data can be reduced, a label data set is expanded by utilizing label-free data during training, network convergence is accelerated, and the classification accuracy of the generative adversarial network is improved, so that the precision of network image classification under a small sample condition is further improved.

Description

technical field [0001] The invention belongs to the technical field of image processing, and in particular relates to a semi-supervised image classification method based on joint training to generate an adversarial network. Background technique [0002] As one of the most common tasks in the field of computer vision, image classification extracts the features of the original image and classifies them according to the features. Traditional feature extraction is mainly achieved by analyzing and processing the color, texture, and local features of the image, such as the scale-invariant feature transformation method, the directional gradient method, and the local binary method. However, these features are all artificially designed, largely relying on human prior knowledge of the recognition target to design, which has certain limitations. With the advent of the era of big data, the image classification method based on deep learning has the ability to process and represent a lar...

AI Technical Summary

Problems solved by technology

However, the GAN network still has problems of poor stability and dependence on label data, so it cannot be directly applied to classification tasks.

[0004] Aiming at the problem of poor stability of GAN network, there are already many methods to solve it by improving GAN network structure or optimization algorithm

However, at present, there is no effective classification method for the problem of relying on label data, so there is an urgent need for an improved GAN network that reduces the network's dependence on label data to a certain extent and can improve the accuracy of network classification.

Images