Predição Intra-Quadro Baseada em Aprendizado Profundo para Light Fields Densos

This study proposes a new strategy for intra prediction of dense light fields by reinterpreting the problem as image inpaiting and using convolutional neural networks. Multiple architectures and training techniques were evaluated in order to identify the most efficient configuration for performing intra prediction in a video encoder for this type of data. Separate networks were trained for each of the 3 block sizes of the encoder and their performance evaluated separately and together. The results showed that the use of convolutional neural networks as intra predictors significantly improves coding efficiency in the EVC encoder, achieving an average BD-rate reduction of -30.53%.