Dichromatic Model Based Highlight Removal via Deep Learning

Chan-Ho Lee, Jun-Sang Yoo, Jong-Ok Kim

Abstract: The dichromatic model has been popularly used for highlight removal. In this paper, we propose a deep learning network which estimates the dichromatic parameters to achieve the highlight removal. Unlike recent works, we utilize not only spatial but also temporal characteristic. The network structure consists of two subnetworks, and the diffuse dictionary and coefficient matrix are generated through each subnetwork. This model is jointly learned by setting 3 loss functions, which split the diffuse and specular components well. To evaluate the performance of highlight removal, we have the diffuse image that highlight is thoroughly removed by calculating the multiplication with the diffuse dictionary and the diffuse weight in a coefficient matrix. This result shows estimating the dichromatic parameters is well implemented in the proposed deep learning method.