Sequential Enhancement for Compressed Video Using Deep Convolutional Generative Adversarial Network

Compression artifacts cause negative visual perception and are tough to reduce because of the balance between compressibility and fidelity. Despite extensive research on traditional methods, they take insufficient effect on quality enhancement. Researches concerning the problem turn to concentrate on quality elevation of single frame using CNNs but ignore the continuity, which is called inter-frame correlation that is critical for video enhancement. There are some CNN-based approaches pursuing good effects, however, sacrificing efficiency. Considering the demand for video quality enhancement and the feature of consecutive frames, this paper proposes a bi-frame generative adversarial network. It takes advantage of inter-frame correlation for bi-frame motion compensation, producing accurate compensated frames. Then, a multi-scale convolutional layer with dilated filters, which constrains parameters and overcomes block effects, is proposed to promote efficiency. Subsequently, a multi-layer deep fusion section is employed to avoid gradients vanishing and realize deep compression artifacts reduction. The ability of discrimination is enhanced with the engagement of a devised relativistic average discriminator which optimizes the whole network. As experiment results demonstrated, bi-frame generative adversarial network shows its effectiveness in terms of various indices. It also presents satisfactory visual performance with comparative test speed compared to listed approaches.