Super-resolution-based enhancement for real-time ultra-low-bit-rate video coding

This paper presents a software-only, real-time video coder/decoder (codec) with super-resolution-based enhancement for ultra-low-bit-rate compression. The codec incorporates a modified JPEG2000 core and interframe predictive coding, and can operate with network bandwidths as low as 500 bits/second. Highly compressed video exhibits severe coding artifacts that degrade visual quality. To lower the level of noise and retain the sharpness of the video frames, we build on our previous work in super-resolution-based video enhancement and propose a new version that is suitable for real-time video coding systems. The adopted super-resolution-based enhancement uses a constrained set of motion vectors that is computed from the original (uncompressed) video at the encoder. Artificial motion is also added to the difference frame to maximize the enhancement performance. The encoder can transmit either the full set of motion vectors or the constrained set of motion vectors depending upon the available bandwidth. At the decoder, each pixel of the decoded frame is assigned to a motion vector from the constrained motion vector set. L ₂-norm minimization super-resolution is then applied to the decoded frame set (previous frame, current frame, and next frame). A selective motion estimation scheme is proposed to prevent ghosting, which otherwise would result from the superresolution enhancement when the motion estimation fails to find appropriate motion vectors. Results using the proposed system demonstrate significant improvements in the quantitative and visual quality of the coded video sequences.