Visual quality optimization techniques applied to rate-control in SVC
In digital video coding, the main aim of bit-rate control is to provide a bit-stream with a final data rate as close as possible to a predefined target one, by simply adapting the quantization parameter to the video sequence complexity. However, in case of low bit-rate applications or highly complex sequences, the bit-rate control would impose a too high quantization step thus inevitably compromising the visual quality of coded pictures. In the depicted case, it is convenient the application of a Frame Skipping strategy. The Frame Skipping technique adaptively skips whole coded pictures, so the encoder could allocate more bits for other frames so as to perform higher image quality and reduce rate criticisms without losing bits. Moreover, the human visual system is particularly sensitive to quantization artefacts and frame rate discontinuity, therefore skipping frame selection must balance quality and rate to avoid visual fluctuation of image quality. This thesis work implies:
- Study and documentation on frame skipping techniques in literature.
- Development and integration of the Frame Skipping technique into the STMicroelectronics proprietary Constant Bit-Rate (CBR) control algorithm. The STMicroelectronics encoder is a C++ software model of the new MPEG SVC (Scalable Video Coding) standard. The implementation will be carried out using GNU C/C++ program languages and will be executed in a general purpose computational platform (PC).
- Finally, extensive objective and subjective evaluations must be done to test the algorithm performance.
The new video coding standard ITU-T/MPEG Scalable Video Coding (SVC) extends the H.264/AVC functionalities with new effective improvements particularly devoted to layered coding in terms of temporal, spatial and quality scalability. A scalable bitstream can contain different representations of the same video sequence, with different temporal or spatial resolutions and quality granularity, all encapsulated according to the standard coding hierarchy. In particular, with the MGS (Medium Grain Scalability) each picture can have multiple quality representations, each one with an increasing associated bit-rate, thus providing the possibility to select different compromises between quality and rate. In this way, SVC enables the transmission and decoding of partial bit-streams to adapt instantaneously to channel conditions, for example in case of video streaming applications as video conferencing services.
The final aim of this thesis work is the study, development and performance evaluation of a transmission system based on the SVC standard and devoted to video streaming applications on variable bandwidth channels. The activity implies:
- outline a set of significant encoding scenarios to be tested;
- channel modelling and simulation;
- extensive objective and subjective evaluations on decoder side of algorithm performance.
Noise reduction is a fundamental aspect for every digital video process since it produces two main benefits: it improves human visual perception and increases compression efficiency of digital video coding systems. Today the most diffused video standard is the H.264/AVC (Advanced Video Coding), which implements a hybrid coding approach by exploiting both temporal and spatial redundancies to achieve efficient compression ratios. The noise, actually, causes high frequency spectral contributions in both spatial and temporal domains, thus preventing the coding process from fully exploitation of temporal and spatial samples correlation. The main targets of this thesis work are:
- noise model shaping of most diffused CMOS sensors for digital cameras and camcorders;
- development of a noise reduction method combining spatial and temporal filtering, also known as MCTF (Motion Compensated Temporal Filtering), employing STMicroelectronics proprietary algorithm implementation;
- integration of the de-noising model as spatio-temporal prefiltering stage of HDTV resolution sequences, according to the H.264/AVC coding standard;
- extensive objective and subjective evaluation of algorithm performance.