Specular highlight removal using a divide-and-conquer multi-resolution deep network

The highlights in the digital images, which are from specular reflections on the surface of the object, are inevitable in the real world. It leads to erroneous results in many image processing algorithms. A specular highlight removal model based on a divide-and-conquer multi-resolution deep network is proposed according to the relation between the image patch and its corresponding maximum diffuse chromaticity. We use a Laplacian pyramid to decompose the image patch into two levels, including the same-sized patch with high- frequency components and the low-resolution blurred patch. The former exploits image textures and preserves local structures, and the latter preserves local intensity. Then we design different sub-networks to extract features for the two levels in a divide-and-conquer way and then fuse the features to achieve each pixel’s maximum diffuse chromaticity. Unlike the state-of-the-art methods where the model is trained in an image space, the proposed model for highlight removal based on the dichromatic reflection model is trained in a patch space. Experimental results demonstrate the proposed model is superior or competitive to the existing methods.