An Improved Intrinsic Image Decomposition Algorithm Based On Retinex And Non-local Texture Constraint

Intrinsic image decomposition remains an long-standing ill-posed problem in the computer vision. In the image, the color information on the surface of the object is determined by a variety of properties, such as light property in the scene, the object’s shape, the reflectance of the object and so on. Barrow and Tenenbaum proposed to use reflectance and shading components to represent these properties. Reflectance describes the reflected characteristic of the object and the shading component represents the light and shade information reflected by the object, which is, the shape of the object surface on the impact of light. In a large number of visual research and Application, If the reflectance information and the brightness information can be separated, And the corresponding visual calculation is made, then better results will be obtained. The intrinsic image decomposition can be widely used in the research of image segmentation, image recognition, texture substitution, and the similarity evaluation between the virtual scene and the real scene. The intrinsic image decomposition problem is that the original image is decomposed into the intrinsic image and the brightness of the original image.This paper improves the condition of non-local texture constraint in this algorithm, which based on Non-local texture constrained Retinex intrinsic image decomposition algorithm that proposed by Li Shen. With the adaptive Kmeans algorithm and the Markov random field theory. It solves the pixel matching error in the original algorithm and the cluster center looking for the wrong problem by the pixels with the same reflectance attributed into a cluster in the whole image. In this way we can get more precise intrinsic figure reflectance and the brightness of the intrinsic figure. At the same time, as the third constraint conditions of the original algorithm of absolute pixel values for the algorithm was improved, we can achieve the ultimate intrinsic image closer to the real intrinsic figure reflectance and the brightness of the intrinsic figure.The experimental results show that the proposed improved in combination with the local texture constrained Retinex intrinsic image decomposition algorithm can get more accurate intrinsic image reflectance and the brightness of the intrinsic images. Intrinsic image algorithm at the same time with the other lateral comparison show that our algorithm can get a higher quality of intrinsic figure reflectance and the brightness of the intrinsic figure, and our result is more close to the MIT provides the standard results.In the end, the paper presents the intrinsic decomposition of image problems in the application of the lane line identification. In this paper, intrinsic decomposition algorithm is applied to track the line identification, when extracts the reflectance of has nothing to do with the lighting information intrinsic image aspirant driveway line recognition in the scene, which is very good solve the traditional lane line identification algorithm under the condition of uneven road lighting information lead to the failure to identify the problem. So it achieves good lane line recognition result.