Modified Multi-scaled Retinex Using Chromaticity of HighlightRegion for Correcting Color Distortion

摘要

As digital still cameras become popular and accordingly theirrnimage quality becomes more of a concern, there are increasingrnstudies in reducing the gap between human-observed scenes andrnimages captured by digital still cameras. The dynamic range of arndigital camera is narrower in contrast to the one of the scene, thusrnit is hard to recognize an object in the shadow region of arncaptured image. The Retinex algorithm is generally used tornimprove detail and local contrast of the shadow region in anrnimage by dividing the image by its local average image, regardedrnas a local illuminant, using a Gaussian filter. The result by retinexrnalgorithm depends on the scale of the Gaussian filter. The smallerrnthe Gaussian filter, the more improved the local contrast, butrnbrings the graying-out and halo artifact. Thus, to reduce thosernartifacts, a multi-scaled retinex algorithm was developed based onrnthe weighted sum of several resulting images from the retinexrnalgorithm by various-scaled Gaussian filters. However, if thernchromatic distribution of the original image is not uniform andrndominated by a certain chromaticity, the chromaticity of the localrnaverage image depends on the dominant chromaticity of thernoriginal image, thereby the colors of the resulting image dividedrnby the local average image are shifted to a complement color tornthe dominant chromaticity of the original image. In this paper, arnmodified multi-scaled retinex method to reduce the influence of therndominant chromaticity in the image is proposed. For this, first, thernlocal average images are divided by the average chromaticityrnvalues of the original image. And then, the local average imagesrnare multiplied by the chromaticity of the global illuminant, whichrnis estimated by the averaged chromaticity values within thernhighlighted region, to consider its influence on the localrnilluminant. In addition, to compensate for the graying-out effect,rnthe chroma value of the output image is enhanced based on that ofrnthe original image in the CIELAB space. Experimental resultsrnshow that the proposed method improved the local contrast andrndetail without color distortion, thereby improving the colorrnrendition.