Novel multi-scale retinex with color restoration on graphics processing unit

Abstract

The multi-scale retinex with color restoration (MSRCR) was developed as a general-purpose image enhancement algorithm that provides simultaneous dynamic range compression, local lightness/contrast enhancement, and good color rendition, and has been successfully used for a wide variety of imagery from diverse fields. While the MSRCR performs good enhancement in most images, the output image can sometimes be further visually optimized during our experiments. An improved MSRCR+Autolevels (AL) algorithm is presented, which can eliminate the impact of a small number of outliers in the histogram of the image and further improve the contrast of an image. New extension significantly improves the visual performance of the MSRCR algorithm. However, the MSRCR+AL containing a large number of complex calculations is computationally expensive, limiting real-time applications. In this paper, a parallel application of the MSRCR+AL algorithm on a graphics processing unit (GPU) is presented. For the various configurations in our test, the GPU-accelerated MSRCR+AL shows a scalable speedup as the resolution of an image increases. The up to 45× speedup (1,024 × 1,024) over the single-threaded CPU counterpart shows a promising direction of using the GPU-based MSRCR+AL in large-scale, time-critical applications. We also achieved 17 frames per second in video processing (1,280 × 720).