Bit-Depth Enhancement via Convolutional Neural Network

Nowadays, many monitors are able to display high dynamic range (HDR) images with high bit-depth for each quantized pixel, but most of the existing image and video contents are of low bit-depth. Therefore, bit-depth enhancement (BE) plays a key role in displaying a low bit-depth image in a high bit-depth monitor. Convolutional Neural Networks (CNNs) and Deep Learning (DL) have recently demonstrated impressive performance in generating realistic high-quality synthetic images, semi-supervised classification, and have been extended for video abstraction and so on. But CNNs or any other deep learning algorithm has not yet been applied to expand image bit-depth so far. In this paper, to fill the gap, we propose a novel algorithm to recover the high bit-depth images via deep convolutional neural network. By training the parameters of the neural networks, the model could learn to recover gradual transition areas and avoid the false contour artifacts, which are commonly seen with traditional bit-depth enhancement algorithms. The experimental results show that our proposed method achieves competitive performance compared with existing bit-depth enhancement methods in terms of PSNR and SSIM with greatly suppressed false contours.