Statistics Inspired Hardware Architectures For Image And Video Processing

Conventional digital arithmetic circuits are designed to operate on a specified range of operand magnitudes.The architecture of these circuits is typically developed to improve the area time characteristics and obtain an energy efficient implementation operating at the desired throughput.Moreover, these circuits are required to provide full precision for the full dynamic range of operands and operate at the speed of worst case.Albeit the correctness of results is ensured, the input data statistics are not taken into account. Resultantly, redundant computations are performed since the information content of input data sequence in certain applications, such as image and video processing, is known to be much less than the simple binary descriptions used by the conventional arithmetic circuits. However, run time identification and exploitation of the latent redundancy in computations is non trivial owing to the diverse statistical nature of input data. Although efforts have been made in the past to design low power architectures by exploiting certain patterns in the magnitudes of the operands, few attempts have been made to improve logic area and processing time efficiency by harnessing the statistical properties of the inputs.It seems conducive to design application specific hardware that explicitly incorporates data statistics while computing and consequently saves precious computation cycles and or logic resources which are otherwise wasted in redundant computations.\ud This thesis proposes hardware design approaches that utilize the inherent redundancy in input operands of image and video processing applications in the implementation of arithmetic circuits to achieve most economical tradeoffs between logic resources, processing time and results precision.Specifically, modifications and enhancements in conventional arithmetic hardware design approaches, namely Distributed Arithmetic, Sub-Expression Sharing, Fast FIR parallel filtering and Approximate Processing have been reported to further enhance their efficiency.Conventional Distributed Arithmetic approach to trade-off logic area with processing speed has been modified to include Memoization based Look Up Tables for storage of partial results from past computations.This modification harnesses bit level redundancies in input operands and leads to decrease in processing time on average while requiring a proportionately lesser increase in hardware resource requirements.The proposed approach has been used to implement Color Space Conversion module for incorporation as Instruction Set Architecture enhancement in open-source Intellectual Property Core for OR1200 32-bit processor. Similarly, conventional Sub-Expression Sharing and Fast FIR parallel filtering techniques to implement low complexity hardware structures have been modified to identify low entropy portions of operands for processing with reduced precision.The ensuing low complexity hardware structures depict negligible loss in output precision when used to process low entropy image data while saving precious logic resources in higher proportion.\ud Merits of incorporating input data statistics in hardware design process have been further illustrated through low precision implementation of statistics-inspired circuits for computing Sum of Squared Error and Normalized Cross Correlation as error metric in template matching applications such as Motion Estimation and Disparity Estimation.The proposed low complexity designs process low and high entropy portions of the operands with appropriate logic efforts and perform better than conventional approximate processing circuits which use brute force quantization of operands as well as output results to reduce hardware complexity.The statistics-inspired hardware designs proposed in this thesis either achieve lower processing time using comparable logic resources or lower hardware cost with minimal impact on precision when compared with the conventional approaches.Efficacy of employing signal statistics information in hardware design process has been shown through proofs of logic resource savings in Field Programmable Gate Arrays based implementations and performance in several real-world applications. Up to 70% hardware savings with minimal impact on output precision in the design of Approximate Squarer and 50% reduction in computation times at full precision for Modified Distributed Arithmetic circuits have been achieved through application of the statistics-inspired hardware design approach