Shengsheng Yu

Distance-based weighted prediction for H.264 intra coding

Intra 4times4 is one of the two primary intra prediction types supported in H.264/AVC. For intra 4times4 prediction, eight modes are to capture orientations of images with DC mode to reflect flat regions. Therefore the smoothly-varying regions which are very common in ordinary images can not be well decorrelated. In this paper, to improve decorrelation performance of intra 4times4 prediction, we propose a distance-based weighted prediction (DWP) method to take the place of DC mode. In DWP, weighted prediction is performed with the weighting coefficients inversely relative to the distances between current pixel and its reference pixels. We also present an integral DWP (iDWP) method which is of much lower computational complexity than DWP. Experimental results demonstrate that DWP and iDWP yield average PSNR enhancements of 0.12 dB and 0.1 dB over current H.264 using 4times4 DC prediction respectively.

A novel background-weighted histogram scheme based on foreground saliency for mean-shift tracking

Effective appearance models are one critical factor for robust object tracking. In this paper, we introduce foreground feature saliency concept into the background modelling, and put forward a novel foreground saliency-based background-weighted histogram scheme (FSBWH) for target representation and tracking, which exploits salient features from both foreground and background. We think that background and foreground salient features are both crucial for target representation and tracking. Experimental results show that the proposed FSBWH scheme can improve the robustness and performance of tracker significantly especially in complex occlusions and similar background scenes.

The training of Karhunen---Loève transform matrix and its application for H.264 intra coding

In H.264/AVC, 4 × 4 discrete cosine transform (DCT) is performed on the residual signals after intra prediction for decorrelation. Actually, residual blocks with different prediction modes exhibit different frequency characteristics. Therefore, the fixed transform matrix cannot match the energetic distribution of residual signals very well, which degrades the decorrelation performance. Fortunately, the energetic distributions of residual blocks with the same mode are relatively coincident, which makes it possible to train a universally good Karhunen–Loève transform (KLT) matrix for each mode. In this paper, an optimal frequency matching (OFM) algorithm is proposed to train KLT matrices for residual blocks and nine KLT matrices corresponding to nine prediction modes of 4 × 4 intra blocks are trained. Experimental results show that KLT with trained matrices yields a persistent gain over H.264 using 4 × 4 DCT with an average peak signal-to-noise ratio (PSNR) enhancement of 0.22dB and a maximum enhancement of 0.33dB.

Linear model-based adaptive prediction for video decoding complexity

This paper proposes a novel approach to predict the video decoding computational complexity. The decoding complexity of each frame is found having an approximate linear relationship with the frame length, whereas the motion Information (MI) and the amount of encoded residual coefficients (ERC) are proved to be the two main factors that affect the variation of the model parameters. The changing rule of MI and ERC for different kinds of video contents are investigated, which in turn derives the variation regularity of the model parameters. Then the correlation between the model parameters of neighboring frames is defined as a piecewise function under the constraint of video motion complexity. The derived piecewise function is used to predict the decoding complexity online, and the prediction error is utilized as the feedback to update the model parameters adaptively. Experimental results show that the proposed method can give fairly accurate prediction for the decoding complexity with very low overhead. The average prediction errors for various test sequences are all within 7%.

A Very Low Bit Rate Video Coding Combined with Fast Adaptive Block Size Motion Estimation and Nonuniform Scalar Quantization Multiwavelet Transform

We describe a very low bit rate video coding framework in which motion correlation between successive video frames is exploited in the multiwavelet transform domain. Some complicated techniques, such as spatial prediction in intra coding, adaptive block size motion estimation, more than one previous frames for prediction in inter frames, and content adaptive binary arithmetic coding (CABAC) are used in H.26L standard. The testing results show that H.26L can greatly outperform MPEG-4 ASP in both PSNR and visual quality. However, the encoding of H.26L costs too much time for it is complex to use fast motion search in adaptive block size motion estimation, and CABAC needs much time to generate the code list for entropy coding. Whereas, only four types of symbol are generated after zero tree wavelet coding so that the entropy coding can cost less time than CABAC. Moreover, if we select 8× 8 sized block as a basic mode, which can be united into the large size mode if neighbored 8× 8 sized blocks have same reference frame and motion vector, then the fast motion estimation can be feasible. Accordingly, a fast motion search algorithm, multiwavelet transform, and a novel adaptive quantization schemer are applied to the proposed coding frame. Experimental results reveal 0.2–0.5 dB increase in coded PSNR at low bit rates over the state-of-the-art H.26L recommendation, and similar improvements over MPEG-4 at high bit rates, with a considerable improvement in subjective reconstruction quality, while simultaneously supporting a scalable representation.

A hybrid M-channel filter bank and DCT framework for H.264/AVC intra coding

In H.264/AVC, discrete cosine transform (DCT) is performed on the residual blocks after prediction. However, the mismatch between variable block sizes and the fixed transform matrix not only degrades decorrelation performance but also causes severe blocky artifacts inside the blocks. In previous work, M-channel filter bank system (MCFBS) was proposed to overcome these defects. However, the increased percentage of encoding time by using MCFBS is very high, especially for intra coding. More seriously, the constructed M-channel filter bank with floating-point coefficients is an obstacle to hardware implementation. In this work, a hybrid M-channel Filter bank and DCT (HMD) framework is proposed for intra coding. Besides, the integer transform of a newly constructed M-channel filter bank is also implemented for HMD. Experimental results demonstrate that HMD can reduce 64–69% of the complexity of MCFBS with negligible quality degradation.

Integer M-channel Filter Bank Design for H.264/AVC Intra Coding

In the previous work, M-channel filter bank systems(MCFBS) were proposed to overcome the mismatch between the fixed transform matrix and variable transform block sizes in H.264/AVC intra coding. However, the constructed filter banks for MCFBS are with irrational coefficients, which bring great obstacles to implement integer transform. In this work, anew class of M-channel filter banks is firstly constructed and then several integer M-channel filter banks (IMFB) derived from this class are proposed for MCFBS. Experimental results show that IMFB can save about 61% of encoding time per frame, with nearly negligible loss of peak signal-to-noise(PSNR) value.

An adaptive packet transmission model for real-time embedded network streaming server

Embedded systems are poised to enable a variety of existing and emerging multimedia streaming applications. Digital signal processor (DSP), with its inherent features of programmable, reconfigurable and efficient, stands a dominate role in such area. However, limited resource is one of the most challenge problems in the DSP-based system. Delay, packet loss and data corruption are inevitable. In this paper, we proposed a content-based feedback control mechanism (CFCM) to operate packets delivery proceeding. Further more, to seek for accurate control and less delay, we employed an adaptive packet transmission mechanism called forward control algorithms (FCA) with using a model which represents the relation among packet transmission interval, link status and buffer capacity. Both CFCM and FCA have been integrated in real-time network video surveillance system to provide efficient and smooth packet transmission. In addition, the system payload (eg. client numbers) is doubled for strict real-time streaming by using FCA.

Adaptive winner-update search for efficient motion vector estimation in video coding

A kind of fast block-matching algorithms which do not introduce any prediction error as compared with the full-search algorithm is to reduce the number of necessary matching evaluations for every searching point in the search window based on partial distortion elimination (PDE), among which winner-update search (WUS) algorithm is a well-known technique. Based on correlation characteristic between gradient magnitudes and matching errors of pixels, this paper proposes an adaptive WUS (AWUS) algorithm which significantly improves the computation efficiency of the original WUS algorithm. This approach can eliminate inappropriate motion vectors faster using an efficient lower bounds set computed according to the edge features of the predicted image. Experimental results indicate that it can accelerate search process by more than 10%, compared with the original WUS algorithm without any loss of estimation accuracy. Furthermore, the proposed algorithm is most suitable for estimating motion vectors for video sequences with rich texture information.

Client error concealment of RAID

A new technique to recover the information loss in a block-based image coding system is developed in this paper. Disk arrays organize multiple, independent disks into a large, high-performance logical disk. However, with more devices, reliability drops. A single disk failure in RAID level 5 will lead to the increase of the load of each surviving disk by 100% for data reconstruction. Each disk is loaded at less than 50% of its capacity in the fault-free state so that the surviving disks will not saturate when failure occurs. By image partitioning, decoder reconstructs DCT blocks in the sub-image and does not impose any significant load on the disk array. Our approach improves the quality of the compressed image by 2dB to 10dB and above and reduces the code overhead of the compressed image data by 2 percent to 10 percent and above for different images due to recovering main AC coefficients.