Human Head-Shoulder Illumination Transfer

Face illumination transfer is an attracting area; however, current methods ignore the illumination state in hair and shoulder regions. This paper proposes a system for human head-shoulder image illumination transfer, and our method proves better results in the hair and shoulder regions. The system consists of three stages: segmentation, transfer and composition. In the segmentation stage, we segment the input image into face, hair, and shoulder. From a head-shoulder images dataset, the system selects best match regions for input image according to their illumination, geometry and reflectance attributes. In the transfer stage, we adopt inhomogeneous illumination transfer strategies for each of the three regions. Finally, all three regions are composited together. Experimental results demonstrate that the inhomogeneous transfer strategy of our system performs better in preserving the input head-shoulder image's identification structure, reflectance attributes and texture detail, and retains the illumination of the original reference better than the homogeneous transfer strategy.