Sparse Coded Decomposition for Single Image-based Specular Removal

In this paper, we propose a method to separate diffuse and specular images from a single image based on sparse representation. The proposed method relax the recent sparse-based technique for removing specular highlights that require multiple input images. Instead, we use only a single input image and then generate multiple specular-free images from this image. The specular removal problem is formulated as the recovery of a sparse representation and an outlier representation. From the obtained specular-free images, a dictionary is reconstructed as the basis function for sparse representation. The sparse signal and the dictionary is used to estimate diffuse image while the specular image is obtained from the outlier. The sparse coding was solved using alternating direction method of multiplier. In addition, a new variant of the specular-free image based on XYZ color space is also proposed to improve sparse representation. The experimental results showed that this method outperformed the accuracy of standard specular removal methods.