Jointly Optimized Transform Coding and Spatial Prediction for Video/Image Compression


Fig.1 Perceptual quality comparison: The proposed hybrid transform coding scheme (top right) significantly reduces blocking effect and preserves more details, in comparison with DCT (bottom left) and DCT w/ deblocking Filter (bottom right).

A novel approach to jointly optimize spatial context-adaptive prediction (employed as intra mode in H.264) and the choice of the subsequent transform coding is developed for video/image compression. The derived optimal transform for the prediction residual is shown to be a close relative of discrete sine transform, with basis vectors that tend to vanish at the known boundary and maximize energy at the unknown side. 

The overall proposed intra-frame coding scheme, implemented in H.264/AVC framework, switches between derived variant of the sine transform and the conventional discrete cosine transform (DCT), depending on the prediction direction and boundary information. It is experimentally shown to substantially outperform the standard (H.264) intra coding. Moreover, it effectively reduces blocking artifacts, since the transform now adapts to the boundary conditions and better exploits inter block correlations. The perceptual quality is demonstrated on a typical example in Fig. 1. An integer version of the sine transform is also available for deployment in concurrence with the existing integer DCT in H.264/AVC.
 

An Estimation-Theoretic Approach to Delayed Video Decoding

We propose an estimation-theoretic delayed decoding scheme which leverages quantization and motion information of one or more future frames to refine the reconstruction of the current block. Unlike conventional standard decoders, which reconstruct blocks immediately as the corresponding quantization indices (as well as motion reference to a prior frame) are available, the proposed scheme offers optimal trading off some decoding delay for improved reconstruction.  The method is implemented in the transform domain; and efficiently combines all available (including future) information in an appropriately derived conditional probability density function, so as to obtain the optimal delayed reconstruction of each transform coefficient in the current block. A complementary fast inverse motion mapping is also developed to construct the ''motion trajectory'' which is central to transform domain implementation (Fig. 2).
 

Fig.2 The proposed delayed video decoder scheme. The current block (blue in Frame n) is first connected to its prior and future companions (blues in Frame n-1 and n+1 respectively) to construct the motion trajectory. Each transform coefficient of the current block is then reconstructed independently using an estimation-theoretic framework.


Experiments demonstrate substantial gains in terms of the reconstruction quality over the standard H.264 decoder. We emphasize that the proposed delayed decoder only employs information available to the standard decoder, without recourse to side information; hence compatibility with the standard syntax and existing encoders is fully retained.

Optimal Delayed Prediction in Scalable Video Coding

Scalable video coding (SVC) is an important paradigm in many video networking applications. The base-layer is encoded as an autonomous bit-stream to enable coarse reconstruction; and refinement is provided in an enhancement layer that typically employs inter-frame or inter-layer prediction in the spatial (pixel) domain. An estimation-theoretic (ET) approach, that optimally combines all available current base-layer information (captured by quantization intervals in the transform domain) and enhancement-layer information via the motion compensated reference, to achieve optimal prediction at the enhancement-layer, had been developed by our lab. The ET approach significantly improves the prediction quality and substantially outperforms other conventional competitors in terms of compression efficiency. The approach was recently extended to exploit the fact that the base-layer is encoded independent of the enhancement-layers, and explore the potential for gains if prediction at the enhancement-layer is delayed to accumulate and incorporate additional future frame information from the base-layer. This is optimally performed within the ET framework. The proposed technique fully exploits all the available information from the base-layer, including any future frame information, and prior enhancement-layer information. It achieves considerable gains over zero-delay techniques including both standard SVC, and SVC with optimal (zero encoding delay) ET prediction.

Fig.3 Block diagrams of scalable video coder. The ET with zero encoding delay (left) optimally combines information from past enhancement layer and current base layer. In addition to the casual information, the proposed ET approach with one frame delay at enhancement layer prediction further utilizes future information from base layer.

Coming soon

  • Spectral coefficient-wise optimal estimate (SCORE) of end-to-end distortion: a dual technique of ROPE in transform domain.
  • Transform domain motion-compensated prediction: exploiting the true temporal correlation.

For more information, please contact us at jingning at umail ucsb edu.