Learning a Code-Space Predictor by Exploiting Intra-Image-Dependencies

We introduce a neural network based image coding model that utilizes a code-space predictor to reduce code length by modelling dependencies within the code. Inspired by the prediction mechanism of inpainting, we learn a spatial predictor in the code space to efficiently deal with spatial dependencies. It is jointly trained with the codec to estimate a probability distribution from adjacent code symbols. The resulting code stores information related to the image and the prediction of neighbours. To improve its optimization, we adapt the Generalized Divisive Normalization into a sparse variant. The resulting model outperforms other prediction based methods. We show that when integrated with the recently proposed hyperprior model, our approach obtains state-of-the-art performance for CNN-based image codecs on the MS-SSIM scale.