IMAGE CLASSIFICATION METHOD FOR IMPROVEMENT OF AUXILIARY CLASSIFIER GAN

摘要

The present invention discloses an image classification method for improvement of auxiliary classifier GAN. The method comprises: changing partial convolutional layers of an discriminator into pooling layers based on ACGAN network structure, matching the feature output of generation specimen in the discriminator with that of real specimen in the discriminator, connecting a Softmax classifier at the output layer of the discriminator network, and outputting the estimated value of posterior probability of a specimen tag; regarding the real specimen as supervision data with tags and the generation specimen as fake data with tags, using the real/fake attribute of the specimen and the entropy loss function of output tags and input tags of the specimen to reconstruct loss function of the generator and the discriminator. Compared with the original ACGAN method and the convolutional neural network having a network structure of the same depth, the described method has better classification accuracy.