An Research On Low-visibility Color Image Enhancement Based On Retinex

Color image, which contains a lot of color, can provide more information than gray image. It is the certain ability of human vision system to perceive color and will not change according to light conditions. However, it is not easy to achieve high quality image due to the reasons such as equipment, illumination and imaging principle. The image will be low-visibility and color distortion so that the use value of it will be affected. Therefore, it has very practical significant of low visibility of color image enhancement, which can be widely use in areas of military, medicine science, civil and public security.To enhance the color image, first we need to know the basic information about color image, including the common color space and the sense principle for human vision to color image. There are common color space including RGB color space, YCbCr color space and HSI color space. These space models and channel meaning are different from each other and each has advantages. Eyes can perceive lightness by columnar cells and perceive color by pyramid cells on retina. Human eyes, which can get rid of the effect of incident light to object spontaneously and estimate the color of object constantly. That is to say, regardless of the change of light conditions or other external factors, the human eyes’color perception is fixed on the same object. Based on this character of human eyes, Land proposed Retinex model.The Retinex model refers to human perception of lightness and color. It has explained both how the human visual system regulating perceived object color and brightness and why the same object’s color is constant under different light source. Based on the Retienx model, Single-scale Retinex, Multi-scale Retinex and Multi-scale Retinex with Color Restoration algorithm are proposed. The multi-scale Retinex(MSR) algorithm, which is high dynamic range, color constancy and high color fidelity, is widely used in low-light image enhancement.This paper improves the conventional retinex algorithm combined with image fusion and image segmentation. First, we propose a novel multi-scale retinex method based on sub-band weighting fusion for image enhancement. Firstly, hybrid intensity transfer function is used to have retinex outputs in different scales. Then the retinex outputs are decomposed into non-overlapping spectral sub-bands. Image enhancement is processed in each sub-band. Finally, the resulted image is fused by weighting every point of each Retinex outputs to emphasize the target information and weak the disturbing information. To avoid severe change, we decomposed the image series to be fused by laplacian pyramid and made use of seamless blend technology to achieve output image.Second, this paper proposes a retinex enhance method based on image divide. First we find the low-luminance areas which need to enhance by image divide algorithm. These areas are enhanced while the other areas keep the same. Therefore the excessive enhancement can be avoided. To refuse the situation that severe changes appear in the edge of the two areas, seamless blend technology of laplacian pyramid can be used. As the retinex enhance algorithm based on image divide mainly divide the image according to the gray value of the pixels, we proposed the method in YCbCr color space, which project the RGB image to YCbCr color space and only enhance the Y channel.Finally, this paper combines the subject and object evaluation to evaluate the quality of the enhanced image. Take the luminance, contrast, entropy and saturation as the object evaluation and the eyes’feeling as the subject evaluation. This paper compared with the traditional enhancement method such as histogram equalization and its improvement, Multi-scale retinex and its improvement et al. The experimental results show that the proposed algorithm effectively enhances the luminance of original image and achieve a visual-pleasing and color-vivid output.