High Bit-Depth Image Acquisition Framework Using Embedded Quantization Bias

In this paper, we present a novel image acquisition framework capable of reconstructing high bit-depth images using an array of low bit-depth scalar quantizers. Our key contribution is a codesign of pixel quantization and reconstruction in image acquisition pipeline. Different from the traditional image acquisition scheme, where each pixel value is quantized and reconstructed independently, the proposed framework is designed based on the interpixel correlations in local image regions. The interpixel correlations imply that quantized pixel values of adjacent locations can be interpreted as multiple descriptions of a common pixel value. Because combining multiple descriptions leads to reduced uncertainty, a pixel value with higher bit depth can be reconstructed by exploiting the interpixel correlations. In this paper, we propose to inject an embedded quantization bias (EQB) to the prequantized image signal and feed the sum to scalar quantizers. Injecting EQB has the same effect as shifting the quantizers by different amounts at different pixel locations, driving the quantized values of adjacent pixels to form informative descriptions of a common pixel value. We present comprehensive studies on this framework, including the optimal design of the EQB signal, the reconstruction strategies for noisy input, and generalized assumptions on interpixel correlations. We show in the experiments that the proposed image acquisition framework significantly outperforms the competing methods and effectively improves the quality of the reconstructed images.

I. Introduction

For efficient representation and processing, a continuous-domain continuous-amplitude image signal is converted to discrete-domain discrete-amplitude image signal in digital cameras [1]. The fixed array of image sensors accounts for the discretization in spatial domain (i.e. sampling), while the discretization in amplitude is achieved by pixel-domain scalar quantization. A scalar quantizer of bit-depth maps a pixel value to one of discrete amplitude levels, so a high bit-depth (HBD) image becomes low bit-depth (LBD) after quantization. This process — the scalar quantization of pixel values— inevitably introduces distortions to the resultant LBD image. A well-known rule in scalar quantization is the “6dB per bit rule” [2], meaning an increment of 1 in bit-depth approximately leads to 6dB's gain in peak signal-to-noise ratio (PSNR) of the quantized signal.