Innovative hole filling method for depth-image-based-rendering (DIBR) based on context learning

摘要

A new convolutional neural network is proposed for hole filling in the synthesized virtual view generated by depth image-based rendering (DIBR). A context encoder in the network is trained to make predictions of the hole region based on the rendered virtual view, with an adversarial discriminator reducing the errors and producing sharper and more precise result. A texture network in the end of the framework extracts the style of the image and achieves a natural output which is closer to reality. The experiment results demonstrate both subjectively and objectively that the proposed method obtain better 3D video quality compared to previous methods. The average peak signal-to-noise ratio (PSNR) increases by 0.36 dB.