New Framework for Highlights Removal Based on CycleGAN

Highlights detection and removal is hot topic in computer vision. It exists in high-gloss leather, glass, plastic, metal parts, and other mirror-reflective objects. Deep convolutional neural network-based highlight removal techniques can learn material rendering parameters supervised by a large number of paired datasets to automatically remove highlights from object surfaces. However, in reality, the datasets are usually unpaired. It is difficult to obtain a large number of paired images. To deal with the unpaired data, dual learning (CycleGAN) can make full use of unlabeled data. Therefore, based on S2D-NET, a new framework is proposed to address the above problem. The main contribution is that: a confidence map with independent averages is proposed to guide the network to quickly search for the initial value, which can solve the problem of slow network convergence due to the lack of a very strict mathematical definition to distinguish specular reflection components from diffuse reflection components. The experiments are compared with the latest three methods. It is found that the SSIM and PSNR values of our proposed algorithm are improved.