BitNet: Learning-Based Bit-Depth Expansion

Abstract

Bit-depth is the number of bits for each color channel of a pixel in an image. Although many modern displays support unprecedented higher bit-depth to show more realistic and natural colors with a high dynamic range, most media sources are still in bit-depth of 8 or lower. Since insufficient bit-depth may generate annoying false contours or lose detailed visual appearance, bit-depth expansion (BDE) from low bit-depth (LBD) images to high bit-depth (HBD) images becomes more and more important. In this paper, we adopt a learning-based approach for BDE and propose a novel CNN-based bit-depth expansion network (BitNet) that can effectively remove false contours and restore visual details at the same time. We have carefully designed our BitNet based on an encoder-decoder architecture with dilated convolutions and a novel multi-scale feature integration. We have performed various experiments with four different datasets including MIT-Adobe FiveK, Kodak, ESPL v2, and TESTIMAGES, and our proposed BitNet has achieved state-ofthe-art performance in terms of PSNR and SSIM among other existing BDE methods and famous CNN-based image processing networks. Unlike previous methods that separately process each color channel, we treat all RGB channels at once and have greatly improved color restoration. In addition, our network has shown the fastest computational speed in near real-time.

A Little Bit More: Bitplane-Wise Bit-Depth Recovery

“…Recently, deep learning-based approaches have gained in popularity [10], [11], [12], [13], [14], [15], [16]. Works by Hou and Qiu [14] and GG-DCNN [15] employ a U-Net style architecture to predict the HBD image given the LBD image.…”

“…The networks of BE-CNN [10] and Liu et al [16] comprise a chain of deconvolution layers with long and short skip connections. BitNet [12] uses an encoder-decoder architecture with dilated convolutions and multi-scale feature integration. BE-CALF [11] employs a chain of convolutionaldeconvolutional layers with dense concatetations of all level features.…”

“…The green box denotes a zoomed-in region of the image. While (q, N )= (8,10) or (8,12) are common target use cases in HBD displays or photo editing applications, most printers and monitors (and even this paper's PDF format) are designed for 8-bit.…”

“…Several classical methods (e.g., [4], [5], [6], [7], [8], [9]) for bit-depth recovery exist in the literature. More recently, deep learning algorithms [10], [11], [12], [13], [14], [15], [16] have been proposed. The strategy employed by most deep learning methods (e.g., [10], [12], [13], [14], [15], [16]) is to train the network using the quantized low-bit-depth (LBD) image as input and have the network predict the high-bit-depth image as output.…”

“…More recently, deep learning algorithms [10], [11], [12], [13], [14], [15], [16] have been proposed. The strategy employed by most deep learning methods (e.g., [10], [12], [13], [14], [15], [16]) is to train the network using the quantized low-bit-depth (LBD) image as input and have the network predict the high-bit-depth image as output. An alternative strategy recently proposed by Liu et al [11] is to train the network to predict the residual between the ground truth HBD image and the quantized LBD image.…”