Adaptive rate control and fast motion vector refinement for heterogeneous transcoding in DCT domain

Abstract

Video communication requires the seamless delivery of video content to a broad range of users with different bandwidth and resource constraints. Transcoding is an essential technique for real-time video communications, video streaming, and VoD (Video on Demand) services over heterogeneous networks which may require different QoS (Quality of Service). In addition to the heterogeneity of present communication networks, the heterogeneity of the user preference for visual quality and picture resolution makes it even more important to design efficient transcoding algorithms. Furthermore, the video compression standards used by the source signal and the client device may not be the same. A possible scenario involves the MPEG-1 and MPEG-4 standards. MPEG-1 was designed for digital storage and CD-ROM application, while MPEG-4 has been proposed for the transmission of video in a wireless and error-prone channel due to its efficient coding scheme and robustness to channel errors. This scenario is becoming particularly important for transmitting stored video over low bandwidth channels such as the mobile networks and the Internet. So far, however, homogeneous transcoding of MPEG-1 to MPEG-1, MPEG-2 to MPEG-2, H.261 to H.261, and H.263 to 1-1.263 have been investigated intensively. Therefore, efficient heterogeneous transcoding algorithms must be designed to fix those heterogeneities. In this dissertation, we propose a heterogeneous transcoding scheme that converts an MPEG-1 bitstream of high bit-rate and spatial resolution into an MPEG-4 bitstream with lower bit-rate and lower spatial resolution in DCT domain. The first technique is an efficient Laplacian model-based requantization for INTRA pictures particularly to improve the picture quality and coding efficiency of INTRA pictures. The quantizer for MPEG-1 INTRA MB usually uses a quantization matrix while the quantizer for MPEG-4 simple profile does not. As a result, the quantization step sizes of the two quantizers may...