基于块匹配与3D滤波的多帧RAW数据低光成像去噪与稳定技术

需积分: 12 47 浏览量更新于2024-09-09 收藏 7.9MB PDF 举报

本文探讨的是多帧原始数据图像的联合去噪问题，特别是在数字成像传感器中遇到的低光环境或快速快门速度下的场景。研究者关注的是利用最新的图像模型来处理这个问题，该模型考虑了信号依赖性噪声和由于传感器曝光不足或过度导致的数据剪切。他们提出的去噪方法是基于V-BM3D算法（Visual-Bilateral Multi-resolution 3D）[5]。V-BM3D算法以其空间-时间的非局部特性而闻名，这使得它在处理原始数据时无需进行精确的帧对齐，从而简化了实际应用。这种方法对于处理高噪声条件特别有效，因为它能够充分利用数据中的冗余信息，通过块匹配和3D滤波技术进行联合处理。预处理和后处理步骤是关键环节，包括了变异性稳态化、偏移校正以及去饱和（即去剪切）等手段。这些变换都是为了确保在处理过程中保持信号的准确性和稳定性，同时减少因曝光异常带来的数据失真。通过这种组合，算法能够在不损害图像细节的前提下，显著降低噪声的影响，提升图像质量，使之适用于低光照环境下，如夜视或者运动物体成像。这篇文章介绍了一种结合了先进图像模型与V-BM3D算法的多帧原始数据去噪策略，旨在解决现代数码相机在极端照明条件下获取清晰图像的技术挑战。其独特之处在于它的鲁棒性和实用性，尤其对于那些对图像质量有极高要求的应用领域。

MULTIFRAME RAW-DATA DENOISING BASED ON BLOCK-MATCHING AND

3-D FILTERING FOR LOW-LIGHT IMAGING AND STABILIZATION

Giacomo Bor acchi

∗

and Alessandro Foi

∗∗

∗

Dipartimento di Elettronica e Informazione, Politecnico di Milano

via Ponzio 34/5, 20133, Milano, Italy

web: http://home.dei.polimi.it/boracchi email: Þrstname.lastname@polimi.it

∗∗

Department of Signal Processing, Tampere University of Technology

P.O. Box 553, 33101, Tampere, Finland

web: http://www.cs.tut.Þ/~foi email: Þrstname.lastname@tut.Þ

ABSTRACT

We consider the problem of the joint denoising of a number of raw-

data images from a digital imaging sensor. In particular, we exploit

a recently proposed image modeling [8] that incorporates both the

signal-dependent nature of noise and the clipping of the data due to

under- or over-exposure of the sensor.

Our denoising approach is based on the V-BM3D algorithm [5],

coupled with a set of homomorphic pre- and post-processing trans-

formations derived for variance-stabilization, debiasing, and declip-

ping [6]. The spatio-temporal nonlocality of V-BM3D frees us from

the need of an expli cit registration of the frames. It results in a prac-

tical algorithm directly applicable to raw-data processing , in partic-

ular for heavy-noise conditions such those encountered in low-light

imaging or imaging at fast shutter speeds.

Experiments with synthetic images and with real raw-data from

CCD sensor show the feasibility of the approach and provide an

indicative measure of the advantage of multiframe versus single-

frame processing.

1. INTRODUCTION

Pictures acquired by digital imaging sensors are al ways subject to

noise. While the signal-to-noise ratio (SNR) can be improved by us-

ing a longer exposure time, this is often not feasible because scene

motion (e.g ., due to movin g objects) or camera motion – also re-

ferred to as camera shake – during the acquisition wo uld result in

blur. The problem is particularly evident when acquiring images at

low-light conditions.

A num ber of diverse solutions have been devised to cope with

this kind of problems. These range from hardware solutions, such

as optical stabilization based on real-time motion-adaptiv e sensor or

lens actuation, to different acquisition paradigms. Particularly ef-

fective for compensating the impact of motion or hand-held camera

shake blur is the approach based on pairs of differently exposed im-

ages [14, 13, 17, 18, 16, 15]. The key idea is to capture two images:

one image taken with a short exposure-time, which ensures that the

blur is ne gligible at the expense of heavy noise, and another image

taken with a longer exposure, which reduces the noisiness but re-

sults in visible blur. Provided som e registration, the noisy image is

used in order to estimate the blur point-spread function (PSF), thus

enabling a non-blind or semi-blind deconv olution of the blurred im-

age. However, scene motion or camera shake v e ry seldom can be

faithfully described as a linear, shift-inv ariant blur; thus, heavy reg-

ularization is necessary to reduce artifacts [17].

This work was supported by the Academy of Finland (application no.

213462, Finnish Programme for Centres of Excellence in Research 2006-

2011, and application no. 118312, Finland Distinguished Professor Pro-

gramme 2007-2010) and by CIMO, the Finnish Centre for International

Mobility (fellowship number TM-07-4952). Part of this work was carried

out during the Þrst author’s visit at Tampere International Center for Signal

Processing (TICSP) in July-October 2007.

An alternative strategy is based on the joint denoising of mul-

tiple images captured sequ entially, thus making the problem con-

ceptually equivalent to a video-denoising problem. In this paper,

we follow this direction and consider the problem in the very spe-

ciÞc setting of raw-data processing, through an observation model

[8] that explicitly incorporates both the signal-dependent nature of

noise and the clipping of the data due to under- or over-exposur e of

the sensor.

In our approach, we rely on the Video Block-Matching 3-D de-

noising algorithm (V-BM3D) [5], coupled with a set of homomor-

phic pre- and post-processing transformations derived for variance-

stabilization, debiasing, and declipping [6]. The spatiotemporal

nonlocality of V-BM3D frees us from the need of an explicit reg-

istration of the frames, while the hom omorphic transformations en-

able an accurate estimation of the true image. Overall, it results

in a practical algorithm directly applicable to multiframe raw-data

processing, which simultaneously extends [5] and [6].

The rest of the paper is organized as follows: Section 2 intro-

duces the observation model and the principal ideas of the V-BM3D

Þlter. The proposed denoising algorithm is detailed in Section 3.

Experiments with synthetic images and with real raw-data from

CCD sensor are presented in Section 4, where we also compare

against the approach based on blurred-noisy image pairs. We con-

clude the paper with few remarks about the impact of redundancy

on the denoising performance.

2. PRELIMINARIES

2.1 Observation model

Let {˜z

}

i=1

, be a sequence set of N raw-data images. According to

[8, 6], each image ˜z

: X →

[

0!1

]

can be modeled as

˜z

=max

{

0!min

{

}}

! x ∈ X ⊂ Z

! (1)

where

= y

,(2)

: X → Y ⊆ R is a deterministic unkno wn original image and

is a zero-mean random error with signal-dependent

standard-de viation $

. Here, $ : R → R

is a deterministic

function while %

is a random variable with unitary variance.

For simplicity, the latter shall be approximated as a standard normal

and all errors are assumed to be independent, thus treating %

i.i.d. with %

∼ N

0!1

. As discussed in [8], this is a suitable

approximation when dealing with the noise in the raw data from

CMOS and CCD digital imaging sensors. For these raw data, the

typical form of the function $ is

=ay

+b! (3)

with the constants a ∈R

and b ∈R depending on the sensor’s spe-

ciÞc characteristics and on the particular acquisition settings (e.g.,

analog gain or ISO value, temperature, pedestal, etc.) [8]. These

two constants are assumed Þxed and invariant during the acquisi-

tion of the images and thus the same for all i =1!&&&! N.

下载后可阅读完整内容，剩余7页未读，立即下载

Captaincafe666

粉丝: 0
资源: 3

基于块匹配与3D滤波的多帧RAW数据低光成像去噪与稳定技术

CVPR 2022 Tutorial Denoising Diffusion-based Generative Model

denoising based on wavelet

a physics-based noise formation model for extreme low-light raw denoising

Image denoising algorithm based on even step-length generalized cross validation model

Note: Parallel processing algorithm of temperature and noise error for micro-electro-mechanical system gyroscope based on variational mode decomposition and augmented nonlinear differentiator

MATLAB Curve Denoising: Removing Impurities and Extracting Useful Signals

【Practical Exercise】Noise Estimation for Various Noises Using MATLAB

【MATLAB Signal Preprocessing】: Data Cleaning and Noise Reduction Methods

【Advanced Chapter】Speech Enhancement and Noise Reduction Techniques in MATLAB

Extracting Useful Information from Noise to Enhance Signal Clarity

最新资源