Combined spatial temporal based In-loop filter for scalable extension of HEVC

Deep learning plays a major role in the present video processing tools and algorithms. To​ alleviate the limitations of present in-loop filters in the scalable extension of HEVC (SHVC), a combined residual network (CResNet) in-loop filtering is proposed in this paper. The proposed CResNet in-loop filter exploits layer information available in the spatial temporal domain to restrain the visual artifacts like blocking and ringing. Particularly, the block information related to current and co-located blocks of the spatial and temporal base layer reference frames are considered to optimize the in-loop filtering. The proposed architecture has four convolution layers at the base layer and two convolution layers at the enhancement layer that significantly reduces the coding complexity and memory. Additionally to completely train the input content and also to enhance the in-loop filter performance, the on/off level control flag for coding tree Unit (CTU) is sensed using rate distortion optimization (RDO) approach. The experimental results demonstrate that the proposed architecture provides up to 6.2% to 7.2% reduction in bit rate and 1.01 dB improvement in PSNR compared to the standard SHVC codec.