Design of unsupervised feature extractor based on InfoGAN with adaptive Info-loss

• 1、School of Information and Communication Engineering, Beijing University of Posts and Telecommunications 100876

Abstract：Unsupervised training of feature vector extractors is an important research field in image processing. The application of effective image feature description vectors in the following tasks, such as image classification and retrieval, can avoid dimension disasters and improve the accuracy of prediction of the tasks. Among many methods, InfoGAN model based on GAN network has great advantages in unsupervised learning to disentangle data features: this model can not only get feature extractors of high-dimensional images through unsupervised training, but also generate images corresponding to different feature vectors by generating module in the model, which can visually distinguish feature categories from the results of generated images. However, this model can only be applied to low-dimensional latent codes to extract the features of the input image, and the smaller the dimension setting, the less the features be got from the image, so the trained model can only describe part of the feature information of the input data, and the output features are unable to be used in the complex image processing tasks, this shortage greatly limits the application range of the model. In this paper, we optimize and improve the information loss function in the original model, and the maximum likelihood is used to replace the original information loss function based on the mean square error in the original model. So, the improved model can achieve stable convergence in the high-dimensional vector, which can disentangle more information from the input data, and be free of the limitation of dimension of the original model to make sure that the model can be used in complex tasks. In the experiment, the improved model and the original model are compared on the MINIST data set. The experimental results show that the original model has training collapse and unable to converge when training with high-dimensional latent codes, but the improved model can converge steadily and output valid image feature expression.

Keywords： information processing technology feature extraction unsupervised learning generative adversarial networks maximum likelihood error