Existing video and image codecs exploit spatial and temporal correlation to achieve compression, but presence of noise reduces this correlation and hence adversely effects their compression efficiency. Noise filtering of the input signal can significantly improve compression efficiency of video encoders. Temporal noise filtering is one such technique that can improve the compression efficiency of encoding noisy content by 5-10% with an AVI reference encoder. Temporal filtering relies on motion information and noise estimation for generating accurate filter coefficients. Generally, encoders have motion estimation modules which can be used for getting motion information, but noise estimation requires a special implementation. In this proposal, we present a low-cost noise estimation method which is used with temporal filtering in AV1 encoding. We calculate the difference between pixels in horizontal and vertical directions to generate the histograms of a difference map. A simple analysis of these histograms provides a very accurate measure of variance of the noise in frames. In comparison to the existing methods, this method has a very low computational overhead.

Macroblock-Tree [1] (also referred as MB-Tree) is a popular and well-established algorithm used to optimally distribute bits across macroblocks within a frame to achieve high compression in video coding. MB-tree’s implementation in x264 [2] encoder brings 10-12% BD-rate PSNR improvement. In x265 [4], a different version of same algorithm is implemented, called CU-Tree. MB-Tree is based on philosophy of assigning more bits (higher quality) to blocks which are most referred in inter frame estimation. This algorithm involves highly computational blocks like motion estimation, intra estimation and cost-propagation. Since this algorithm works in a look-ahead manner (generally as part of rate-control module) to the encoder, it poses a big challenge for real-time implementation like broadcasting and online video transmissions. To reap its benefits in real-time encoding we propose steps to reduce its complexity up to half, without compromising on objective video quality. The results are provided with high performance real-time 4Kp60 H.264 HW encoder inside AMD-Xilinx’sOR Zynq UltraScale+ MPSoC [6].

This paper presents a novel method for protecting privacy sensitive regions inside a digital frame. The purpose is achieved by deteriorating the visual quality of privacy sensitive region by adding probability model based random noise to nonzero quantized transform coefficients of the region. The probabilistic model based noise is generated based on the information from neighboring macroblocks and added only to selective non zero AC transform coefficients to have minimal impact on bit rate. Moreover, the parameters of the probability model are derived from the neighboring blocks inside the frame such that the noise follows characteristics of the frame content. For the purpose of implementation, we used H.264/AVC video coding standard for videos and JPEG image compression standard for images. The probabilistic model based noise is generated using a secret key, which can be shared with the authorized decoder such that the obfuscated region can be completely recovered. In order to limit drifting of obfuscation effect outside the targeted regions, Inter and Intra prediction blocks inside H.264/AVC encoder were modified. Overall, we achieve good obfuscation of selected regions with minimal impact on coding efficiency with the capability to achieve complete reconstruction at the authorized decoder.

In this paper we propose a fast candidate selection method in RMD step of Intra prediction. The proposed method consist of two steps, in the first step, coarser steps between the prediction directions are used as possible candidates for comparison, where the coarser step size is a function of Prediction Unit size. This way we find the dominant direction with minimum cost. In the second step, the dominant direction is refined by checking the prediction directions around the dominant direction to find the best match.

The document, which includes three blank pages, was not presented at the conference and therefore was not made available for publication as part of the conference proceedings.

Improved performance of Intra Prediction in HEVC in comparison to H.264/AVC is primarily due to two main feature enhancements, first is increase in the number of intra prediction directions and secondly due to the support for larger transform sizes. Using lagrangian based rate distortion measure to decide the appropriate prediction direction and Transform Unit size requires calculation of bits generated using complete cabac encoding for every candidate, which is computationally very exhaustive. In order to avoid complex cabac encoding is to use only distortion as a comparison measure, but it has a detrimental effect on coding efficiency. In this paper we present a method based on regression, to estimate the cabac bits for direction mode and for Transform Unit based on the magnitude of non-zero coefficients. This way we avoid calculating cabac encoding for bits calculation and the results show that our method gives 5.4% average BD-rate gain for Luma in comparison to using only distortion as a comparative measure.

This paper describes a method to optimize the power needed to display a video by better illuminating the regions of interest compared to the rest of the video. This method identifies the regions of interest by utilizing parameters from the digital video codecs. It also employs temporal and spatial filters to improve perceptual quality of the video. For our experiment we used MPEG-2 encoded video bitstream and used its motion vectors for identifying the regions of interest, and the results show that the regions of interest are better highlighted in comparison to uniformly scaled illumination for similar power savings. In addition, the method has very low complexity of 0.5 MIPS and low memory footprint of 3 KiloBytes only, so it does not add substantial overhead to the computational complexity of the system as well.

H.264 or AVC (Advanced Video Coding) is a latest digital video codec standard which was developed as an answer to the growing demand for better compression in a wide range of applications and for improved network friendliness. H.264 is able to deliver a compression efficiency of up to 50% over a wide range of bit rates and video resolutions compared to previous standards (e.g. MPEG2 or H.263). The downside is that the H.264 decoder complexity is nearly four times higher than the previous standards. Hence a powerful hardware platform is required to provide real-time performance of H.264 in today's high-end applications like HDTV. We present an innovative Hardware architecture to perform real-time H.264 CABAC decoding using Finite State Machines (FSMs) for decoding of syntax elements. This architecture delivers a throughput of 1 bin per cycle @ 180MHz as reported by Synopsys Design Compiler.

This paper describes an innovative, pipelined, cache-based architecture for a motion estimation coprocessor based on a predictive/recursive algorithm whose computational complexity is low and independent from the search window. The algorithm and the associated architecture yields itself very well to low-power, low-cost video capture devices with low processing capabilities, such as mobile phones, PDAs, or handhelds. The synergy between architecture and algorithmic features allows a high quality output, low memory to cache bandwidth requirements, and a search window independent implementation for H.264/AVC real time video encoding of up to high definition video (HDTV).

This paper describes an innovative algorithm for adding SVC Spatial Scalability to all existing non-scalable H.264/AVC video streams. The algorithmic system builds on a full-decode-full-encode method of conversion and employs reuse of available data by an efficient downscaling of video information for different layers, thus reducing the complexity of the algorithm by manifold whilst maintaining a high coding efficiency. The complexity is further reduced by bypassing of various time consuming processes of the encoding process by the use of sharing and reuse of video information coming from original bitstream. Extensive analysis has been done to evaluate the gains in the coding efficiency and complexity of the algorithm using the full-decode-full-encode approach as a benchmark. The ultimate reduction in complexity is at least 60% from full-decode-full-encode while maintaining the output video quality. The modular nature of the algorithm is designed to be easily adaptable to other scalable video standards as well.

This paper describes a transcoding algorithm able to transcode a compressed video bitstream into a WMV9 (VC1) compliant bitstream. A Closed-Loop Spatial Domain Transcoding approach, modeled on the Cascaded Pixel Domain Transcoder (CPDT) is presented here. It incorporates reusing of information obtained from the input bitstream. To enhance the performance of the transcoder, an approach for pipelining the encoder/decoder processes in the transcoder is proposed here. These modifications can be easily adapted to all the transcoders producing WMV9 as output, the input for whom may include video streams of type MPEG-2, MPEG-4 or H.264. As a proof of concept, the suggested approach has been applied to the case of MPEG-2 to VC1 transcoding. Extensive analysis has been performed and the ultimate gain in transcoding speed achieved for the presented approach compared with MPEG-2 to VC1 CPDT, is in the range of 3 to 7 times faster than CPDT.