New End-to-end Network for Stereo High Dynamic Range Imaging

Cameras people use in their daily life usually can only obtain low dynamic range (LDR) images. In order to obtain high dynamic range (HDR) images, various methods have been invented. But there is a significant problem with most HDR techniques, namely that original HDR methods require images with different exposure conditions to be taken. In this process, if the captured objects are in motion, the generated HDR image will suffer from ghosting artifacts. To solve this problem, one way is to use different cameras to take images with various exposures simultaneously; by this method the impact of object motion can be minimized. Inspired by this idea, we propose MVMEFNet, an end-to-end network that consists of two sub-networks: Warp Net which is used to align the images taken from two views and produce a disparity map, and Fusion Net which is designed to fuse the aligned left view and right view images. We also innovatively introduce deformable encoder in the Fusion Net, which allows for better error correction of the results in warp net. The experimental results show that our proposed method can obtain stereo HDR image with good visual quality.