Pattern Recognition and Image Analysis Computer Vision and Pattern Recognition
Training Normalizing Flow Models in the Latent Space of Variational Autoencoders for Image Entropy Compression
Abstract
Abstract The paper is devoted to the problem of image compression by variational autoencoders (VAEs). One of the drawbacks of the latent space structure of the autoencoder model for the effective entropy compression is the deviation of the empirical distribution of latent vectors from the normal or Laplace distributions, which makes it incompatible with the existing entropy encoders. To improve the properties of the latent space, an approach is proposed that is based on the training of a normalizing flow model for transforming the distribution of latent vectors into standard normal or Laplace distributions. The normalizing flow model is based on an architecture with affine layers and invertible convolutions and is trained on the latent data of pretrained autoencoders. Variational autoencoders used in multimodal generative models and the next-generation neural network codec are taken for experiments. We found that an increase in the arithmetic entropy compression rate significantly depends on the approach to the regularization of the latent space when training the corresponding autoencoders and on the quantization method. The application of normalizing flows for an autoencoder taken from a generative model made it possible to increase the compression rate by 5.2% without loss of the reconstruction visual quality on the standard test set of images. The method of combining autoencoders and normalizing flows for image compression problems provides a basis for new approaches to the optimization of latent representations and increasing the compression efficiency in existing systems without their retraining.