Quality Enhancement of Compressed Video via CNNs

This work presents a compressed video enhancement algorithm based on convolutional neural networks (CNNs), which aims to establish an end-to-end mapping function between the compressed and the original video frames. Finally, the data is stored or transferred in the form of video plus network parameters. We can get a quality improved frame when taking the compressed frame as the input of CNN. Unlike some traditional filtering methods, our method can utilize more prior information, and thus can recover more details. Our algorithm relies on the training data, and independent of the compression method itself. In this work, we adopt H.264/AVC as a typical encoder to verify the effectiveness of our algorithm. Experimental results demonstrate that the proposed algorithm provides better reconstructed quality than that of classical approaches.