Low-light Stereo Image Enhancement Using Convolutional Neural Network

Abstract

We present a method that can increase the quality of a low-light stereo image. While traditional stereo imaging methods have focused on estimating depth from stereo images, our method utilizes stereo images to enhance the low-light condition. The critical challenge for enhancing the low-light condition of stereo images is the disparity between the left and the right images. We proposed an end-to-end convolutional neural network to enhance the low-light condition in stereo images without estimating the disparity. Our proposed network has two sub-networks: the rst network learns how to enhance the low-light condition of stereo images in luminance, and the second network learns how to reconstruct a normal-light full-color image from enhanced luminance and chrominance of the input image. Our two-stage joint network enhances the low-light condition of stereo images significantly more than single-image low-light enhancement method.