Estimation of Bandlimited Grayscale Images From the Single Bit Observations of Pixels Affected by Additive Gaussian Noise

Abhinav Kumar, Animesh KumararXiv 2016 · OpenReview

Abstract

The estimation of grayscale images using their single-bit zero mean Gaussian noise-affected pixels is presented in this paper. The images are assumed to be bandlimited in the Fourier Cosine transform (FCT) domain and oversampled over their Nyquist rate. We propose a non-recursive approach based on first-order approximation of the Cumulative Distribution Function (CDF) to estimate the image from single-bit pixels. The decay rate for mean squared error of estimating such images is found to be independent of the precision of the quantizer and it varies as $O(1/N)$ where $N$ is the "effective" oversampling ratio with respect to the Nyquist rate in the FCT domain.

Introduction

Digital imaging typically relies on multi-bit quantization to represent pixel intensities. However, this paper explores the possibility of reconstructing high-quality grayscale images from extremely limited 1-bit observations—binary "black or white" pixels. This concept hinges on the "quantizer precision indifference principle," which suggests that if a signal is oversampled sufficiently, the precision of the quantizer becomes secondary to the oversampling ratio $N$ in determining reconstruction accuracy.

Methodology

The authors treat the image as a bandlimited signal in the Fourier Cosine Transform (FCT) domain. FCT is chosen for its efficiency in concentrating image energy compared to the standard Fourier Transform. The core of the estimation method relies on three components:

  1. Dither and Noise: The image is corrupted by zero-mean Gaussian noise and a dither signal before quantization.
  2. CDF Linearization: The Cumulative Distribution Function (CDF) of Gaussian noise is naturally S-shaped, but it can be approximated as a linear function within the image's normalized intensity range (typically $[-1, 1]$) if noise variance is sufficiently high.
  3. Non-Recursive Estimation: Using this linearity, the authors developed an algorithm that avoids complex iterations, instead using a Zakai-class interpolation kernel to low-pass filter binary samples back into grayscale estimates.
Extracted figure from page 2 of Estimation of Bandlimited Grayscale Images From the Single Bit Observations of Pixels Affected by Additive Gaussian Noise

Experimental Results

The methodology was tested on both synthetic cosine images and standard benchmarks like the "Peppers" image.

Reconstruction of the 'Peppers' image from 1-bit pixels at different oversampling ratios.

Key findings include:

  • Visual Fidelity: At an oversampling factor of $N=32$, images reconstructed from 1-bit data are visually almost identical to the originals.
  • Mathematical Decay: For perfectly bandlimited synthetic images, the Mean Squared Error (MSE) decreases as $O(1/N^2)$. For real-world images like "Peppers," which are not perfectly bandlimited, the observed distortion varies as $O(1/N)$.
  • Noise as a Feature: Counter-intuitively, the presence of Gaussian noise is beneficial; it acts as a stochastic bridge that allows the 1-bit threshold to capture underlying grayscale values through the linearized CDF.

Conclusion

This research demonstrates that the rate at which quantization error disappears depends primarily on the oversampling ratio $N$, not the sensor's bit-depth. This "precision indifference" suggests that oversampling is a universal remedy for low-bit quantization. Future work may explore extending these principles to other mathematical bases like Wavelets or combining them with Compressed Sensing techniques.

Summary updated

Abstract

The estimation of grayscale images using their single-bit zero mean Gaussian noise-affected pixels is presented in this paper. The images are assumed to be bandlimited in the Fourier Cosine transform (FCT) domain. The images are oversampled over their Nyquist rate in the FCT domain. We propose a non-recursive approach based on first order approximation of Cumulative Distribution Function (CDF) to estimate the image from single bit pixels which itself is based on Banach's contraction theorem. The decay rate for mean squared error of estimating such images is found to be independent of the precision of the quantizer and it varies as $O(1/N)$ where $N$ is the "effective" oversampling ratio with respect to the Nyquist rate in the FCT domain.
BibTeX
@misc{kumar2016estimation,
  author = {Abhinav Kumar and Animesh Kumar},
  title = {{Estimation of Bandlimited Grayscale Images From the Single Bit Observations of Pixels Affected by Additive Gaussian Noise}},
  year = {2016},
  eprint = {1610.08627},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1610.08627},
}
Figure