Improved Retinex Image Enhancement Algorithm

Abstract Retinex method mainly consists of two steps: estimation and normalization of illumination. How to extract the background illumination accurately is a key problem. The backgrounds of picture sequence in video's adjacent frames are usually similar and closely related. More accurate illumination information can be extracted when this characteristics of video's picture sequence is considered. In the paper, we propose an improved Retinex algorithm. Filter the images using the Gauss masks of different scale and parameter for each frame image, and all these filtering results are fused together by minimum method. In the paper, the scale of Gauss filters are set as 5, 9, 13, 25, and their variance set as 0.3, 0.5, 0.7 and 1.0 respectively. 6 adjacent frame images are selected, and the uniform and optical background image for these 6 images can be extracted by maximum method. This method makes use of the similarity and relationship among the adjacent frame images in videos. Enhance the images using Retinex method with this optical background image as their uniform illumination information. Experiment shows that more accurate back grounds are acquired and more excellent enhancement performance are achieved.