Autoencoders with Variable Sized Latent Vector for Image Compression

Learning to compress images is an interesting and challenging task. Autoencoders have long been used to compress images into a code of small but fixed size. As different images need different sized code based on their complexity, we propose an autoencoder architecture with a variable sized latent vector. We propose an attention based model which attends over the image and summarizes it into a small code. This summarization is repeated many times depending on the complexity of the image, producing a new code each time to encode new information so as to get a better reconstruction. These small codes then form sub-units of the final code. Our approach is quality progressive and has flexible quality setting which are desirable properties in compression. We show that the proposed model shows better performance compared to JPEG