Real-Time High-Quality Specular Highlight Removal Using Efficient Pixel Clustering

On the basis of the dichromatic reflection model, recent specular highlight removal techniques typically estimate and cluster illumination chromaticity values to separate diffuse and specular reflection components from a single image. While these techniques are able to obtain visually pleasing results, their clustering algorithms suffer from bad initialization or are too costly to be computed in real time. In this paper, we propose a high-quality pixel clustering approach that allows the removal of specular highlights from a single image in real time. We follow previous work and estimate the minimum and maximum chromaticity values for every pixel. Then, we analyze the distribution pattern of those values in a minimum-maximum chromaticity space to propose an efficient pixel clustering approach. Afterwards, we estimate an intensity ratio for each cluster in order to separate diffuse and specular components. Finally, we present optimization strategies to implement our approach efficiently for both CPU and GPU architectures. Experimental results evaluated in the available dataset show that the proposed approach is not only more accurate, but is also two times faster than the state-of-the-art when running solely on the CPU. Running on the GPU, we show that our approach requires ≈24 milliseconds to remove specular highlights in an image with 3840×2160 (4k) resolution. That makes our GPU implementation more than one order of magnitude (20×) faster than the state-of-the-art for 4k resolution images, while providing the desired effect accurately.