Intra-Frame Coding Using a Conditional Autoencoder

Exploiting spatial redundancy in images is responsible for a large gain the performance of image and video compression. The main tool to achieve this called intra-frame prediction. In most state-of-the-art coders, intra prediction applied block-wise fashion. Up now angular was dominant, providing low-complexity method covering variety content. With deep learning, however, it possible create methods wider range content, being able predict structures which traditional modes can not accurately. Using conditional autoencoder structure, we are train single artificial neural network perform multi-mode paper, derive approach from general formulation intra-prediction problem introduce two extensions mode chroma support. Moreover, propose novel latent-space-based cross component We show power our scheme with visual examples report average gains 1.13% Bjøntegaard delta rate luma 1.21% compared VTM using only modes.