Dark-Channel Enhanced-Compensation Net: An end-to-end inner-reflection compensation method for immersive projection system

Abstract

Immersive projection display system is widely adopted in virtual reality and various exhibition halls. How to maintain high display quality in an immersive projection environment with uneven illumination and the color deviation caused by the inter-reflection of light is still a challenging task. In this paper, we innovatively propose a deep learning-based radiation compensation for an L-shaped projector-camera system. This method employs complex reflection phenomena to simulate the light transport processing in an L-shaped environment, we also designed a Dark-Channel Enhanced-Compensation Net (DECNet) which composed of a convolutional neural network called Compensation Net, a DarkChannelNet and another subnet (such as sensing network) aiming at achieving high-quality reproduction of projected display images. The final output of DECNet is the compensation image to be projected. It is always a critical problem to establish appropriate evaluation and analysis indexes throughout the research of light pollution compensation algorithms. In this paper, PSNR, SSIM, and RMSE are proposed to quantitatively analyze the image quality. The experimental results show that this method has certain advantages in reducing the inter-reflection of the projection plane. And our method could also well replace the traditional process using the backlight transmission matrix. It can be concluded to a certain that this method can be extended to other more complex projection environments with strong scalability and inclusiveness.