Dynamically Reconfigurable Architectures and Systems for Time-varying Image Constraints (DRASTIC) for Image and Video Compression

In the current information booming era, image and video consumption is ubiquitous. The associated image and video coding operations require significant computing resources for both small-scale computing systems as well as over larger network systems. For different scenarios, power, bitrate and image quality can impose significant time-varying constraints. For example, mobile devices (e.g., phones, tablets, laptops, UAVs) come with significant constraints on energy and power. Similarly, computer networks provide time-varying bandwidth that can depend on signal strength (e.g., wireless networks) or network traffic conditions. Alternatively, the users can impose different constraints on image quality based on their interests. Traditional image and video coding systems have focused on rate-distortion optimization. More recently, distortion measures (e.g., PSNR) are being replaced by more sophisticated image quality metrics. However, these systems are based on fixed hardware configurations that provide limited options over power consumption. The use of dynamic partial reconfiguration with Field Programmable Gate Arrays (FPGAs) provides an opportunity to effectively control dynamic power consumption by jointly considering software-hardware configurations. This dissertation extends traditional rate-distortion optimization to rate-quality-power/energy optimization and demonstrates a wide variety of applications in both image and video compression. In each application, a family of Pareto-optimal configurations are developed that allow fine control in the rate-quality-power/energy optimization space. The term Dynamically Reconfiguration Architecture Systems for Time-varying Image Constraints (DRASTIC) is used to describe the derived systems. DRASTIC covers both software-only as well as software-hardware configurations to achieve fine optimization over a set of general modes that include: (i) maximum image quality, (ii) minimum dynamic power/energy, (iii) minimum bitrate, and (iv) typical mode over a set of opposing constraints to guarantee satisfactory performance. In joint software-hardware configurations, DRASTIC provides an effective approach for dynamic power optimization. For software configurations, DRASTIC provides an effective method for energy consumption optimization by controlling processing times. The dissertation provides several applications. First, stochastic methods are given for computing quantization tables that are optimal in the rate-quality space and demonstrated on standard JPEG compression. Second, a DRASTIC implementation of the DCT is used to demonstrate the effectiveness of the approach on motion JPEG. Third, a reconfigurable deblocking filter system is investigated for use in the current H.264/AVC systems. Fourth, the dissertation develops DRASTIC for all 35 intra-prediction modes as well as intra-encoding for the emerging High Efficiency Video Coding standard (HEVC)