Calibrated Relighting Network for Image Light Transfer

• Published in: 2022 4th International Conference on Data Intelligence and Security (ICDIS) pp. 267 - 274
• Main Authors: Li, Cong, Rao, Yuan, Yang, Jian, Yang, Kai, Fan, Hao, Dong, Junyu
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.08.2022
• Subjects: Deep learning; Estimation; Feature extraction; Illumination Estimation; Image Relighting; Light Source Transfer; Lighting; Measurement; Security; Semantics
• DOI: 10.1109/ICDIS55630.2022.00048

Image relighting task aims at changing the lighting conditions while keeping the same contents in the image. Recently, image relighting has witnessed vast progress with the development of deep learning. However, most existing methods solely tackle this problem by treating it as the style transfer task, and lead to irregular illumination and shadows since they ignored complex relations of the coupled illumination and shadows. The latest research presents the decoupled framework for image relighting by decomposing the relighting process into three tasks: scene reconversion, shadow estimation and image re-rendering. In this paper, We follow the decoupled framework and propose a novel Calibrated-DConv block in recovering primary structure and estimating shadow effects by recalibrating the feature and thus enhancing feature representation. Furthermore, we design a novel re-rendering network, which adaptively repaints the image by estimating the weights of interest parts. Experimental results on the VIDIT dataset show that the proposed method performs favorably against recent state-of-the-art methods and achieves reasonable image light source transferring.