基于L0矩阵范数正则化的自然图像去反光算法

去除图像的反光是计算机视觉和计算机图形学的一个基础研究问题。虽然各种方法已经提出,但由于存在丰富的纹理、复杂的背景、遮挡和颜色照明等,去反光的效果并不佳,有待进一步完善。本文针对自然图像的两个观察结果：(1)高光通常是小尺寸和稀疏分布的；(2)剩余的漫反射图像可以用少量基色与稀疏和低秩加权矩阵的线性组合表示,提出一种基于L0范数正则化图像去反光算法。通过L0范数正则化保证编码系数的稀疏性以及核范数保证编码系数的低秩性来恢复那些高光区域的漫反射分量。此外,根据加色混合理论和光照定义,编码系数和高光也分别受到非负性影响。通过对比相关的图像反光去除算法,实验说明了所提出的算法具有较好的优势。

Highlight removal algorithm for real-world images based on L0 matrix norm regularization

Specular highlight removal is a basic research problem in computer vision and computer graphics. Although various methods have been proposed, they generally do not work well for natural images due to rich textures, complex materials, hard shadows, occlusion, and color lighting,so it needs to be further improved. In this paper, a new method of highlight removal for natural images is proposed. Our method is based on two observations of natural images: (1) highlights are usually small in size and sparsely distributed; (2) The remaining diffuse image can be represented by a linear combination of a few primary colors and a sparse and low-rank weighted matrix. Based on these two observations, we design an optimization framework for estimating both diffuse and specular images from a single image. Specifically, we restore the diffuse components of those high-light regions by using L0 norm guarantee sparsity of coding coefficients and nuclear norm to guarantee low rank of coding coefficients. In addition, according to the additive color mixing theory and the illumination definition, the coding coefficient and the specular light are also affected by non-negative effects respectively. The effectiveness and superiority of the proposed method are verified by a lot of experiments on various images.