CSR: Small: Dynamically Reconfigurable Architectures for Time-Varying Image Constraints (DRASTIC) Based on Local Modeling and User Constraint Prediction

University Of New Mexico, Albuquerque NM

Investigators

Abstract

The use of digital video in embedded and communications systems has risen dramatically in recent years. User needs and interests in digital video vary based on video content and available computing resources such as battery life and communications bandwidth. Thus, there is a strong need to develop methods that can dynamically reconfigure hardware and software resources in real-time to respond to changing video content, user needs, or available computing resources. The proposed research will develop methods that will provide run-time management of hardware and software resources for video processing and communications that are jointly optimal in terms of energy, bandwidth, and throughput. Digital video processing and communication often consumes the majority of computing resources and bandwidth. With the emergence of the High-Efficiency Video Coding (HEVC) standard, there is a focus on the development of parallel architecture solutions that can effectively provide real-time coding of high-resolution video while reducing the bandwidth requirements up to 50% from the previous coding standard (H.264/AVC). However, the HEVC's focus on rate-distortion optimization does not consider how computing architectures should adapt to time-varying energy constraints. The proposed research will be focused on the development of Dynamically Reconfigurable Architecture Systems for satisfying Time-varying Image processing Constraints (DRASTIC) that optimize computing resources to satisfy time-varying constraints on energy, bandwidth, and image quality for HEVC and video analysis based on 2D/3D filterbanks. The research is transformative in two different ways: (i) it supports the automatic generation of real-time varying constraints based on video content and available energy while eliminating the need for user inputs, and (ii) it uses a local model to significantly reduce the requirements for estimating large Pareto fronts over a large space of videos. These transformative approaches can significantly expand the applicability of the proposed system and methods.