Real-time retinex image enhancement: algorithm and architecture optimizations

Abstract

The field of digital image processing encompasses the study of algorithms applied to two-dimensional digital images, such as photographs, or three-dimensional signals, such as digital video. Digital image processing algorithms are generally divided into several distinct branches including image analysis, synthesis, segmentation, compression, restoration, and enhancement. One particular image enhancement algorithm that is rapidly gaining widespread acceptance as a near optimal solution for providing good visual representations of scenes is the Retinex. The Retinex algorithm performs a non-linear transform that improves the brightness, contrast and sharpness of an image. It simultaneously provides dynamic range compression, color constancy, and color rendition. It has been successfully applied to still imagery---captured from a wide variety of sources including medical radiometry, forensic investigations, and consumer photography. Many potential users require a real-time implementation of the algorithm. However, prior to this research effort, no real-time version of the algorithm had ever been achieved. In this dissertation, we research and provide solutions to the issues associated with performing real-time Retinex image enhancement. We design, develop, test, and evaluate the algorithm and architecture optimizations that we developed to enable the implementation of the real-time Retinex specifically targeting specialized, embedded digital signal processors (DSPs). This includes optimization and mapping of the algorithm to different DSPs, and configuration of these architectures to support real-time processing. First, we developed and implemented the single-scale monochrome Retinex on a Texas Instruments TMS320C6711 floating-point DSP and attained 21 frames per second (fps) performance. This design was then transferred to the faster TMS320C6713 floating-point DSP and ran at 28 fps. Then we modified our design for the fixed-point TMS320DM642 DSP and achieved an execution rate of 70 fps. Finally, we migrated this design to the fixed-point TMS320C6416 DSP. After making several additional optimizations and exploiting the enhanced architecture of the TMS320C6416, we achieved 108 fps and 20 fps performance for the single-scale, monochrome Retinex and three-scale, color Retinex, respectively. We also applied a version of our real-time Retinex in an Enhanced Vision System. This provides a general basis for using the algorithm in other applications.