Elsevier

Neurocomputing

Volume 188, 5 May 2016, Pages 102-112
Neurocomputing

Comparison of different image denoising algorithms for Chinese calligraphy images

https://doi.org/10.1016/j.neucom.2014.11.106Get rights and content

Abstract

Rubbing is one of the most universal and perhaps the oldest of the techniques that have been used in printmaking. A carefully made rubbing provides an accurate and full-scale facsimile of the surface reproduced. However, many rubbing have been destroyed or lacked a good ways to identify them by certain events, while some other contained a large white background, or have become illegible due to erosion. In order to correct interpretation of these images, some image restoration techniques are employed. Image denoising is one of the important fields in the restoration arena. But, a great challenge of image denoising is how to preserve the edges and all fine details of a rubbing image while reducing the noise. This paper presents a comprehensive comparative study of image denoising techniques relying on Anisotropic Diffusion filter, Wiener filter, TV (Total Variation), NLM (Non-Local Means, NLM), Bilateral filtering. A quantitative measure of comparison is provided by the PSNR, MSE, SNR, UQI and SSIM of the image. Finally, the paper also analyzes its effect of denoising on rubbings with various algorithm and points out the advantages and disadvantages in application.

Introduction

A rubbing is a reproduction of the texture of a surface created by placing a piece of paper or similar material over the subject and then rubbing the paper with something to deposit marks, most commonly charcoal or pencil, but also various forms of blotted and rolled ink, chalk, wax, and many other substances as well [1]. Rubbing is a technique which invented in Wei, Jin, South and North Dynasties (220–589 AD), to make copies of inscribed records, using paper and ink. We can say that, rubbing is a unique way of documentation in China. It has made great contribution to the preservation of the Chinese culture, which is of high historical and aesthetic value and used in many ways in today׳s scientific research, and whose aesthetic feeling cannot be replaced by photographs. The use of rubbing was initially limited to making copies of stone inscriptions and then gradually expanded to bronze ware, jade ware, coins, ink stones, seals, tiles, wood ware, and even to oracle bones, objects dating from the Qin Dynasty (221–207 BCE) to the Ming Dynasty (1368–1644 CE).

The rubbings were made on ancient stone stelae and tomb tablets are not only important carriers of China ancient civilization but also are a classical template to research and learn the art of calligraphy. For the study of the history of writing and calligraphy, from the earliest script on shell and bone down to the running and cursive styles of later masters, inscriptions are irreplaceable sources. They have been tracing the evolution of writing, century after century, also.

Many Chinese ancient calligraphy work created by former famous calligraphers were carved on stone tablets, and calligraphy documents were produced by rubbing. Stone rubbings taken from them have been reproduced and reprinted widely and studied by generations of students, used models to learn and practice the art of calligraphy. Rubbings of engraved models of calligraphy, known as model writing (fa tie) are the most widely reproduced and consulted genre of rubbings in China, Japan, and Korea today.

In recent years, with the development of the scanning equipment and digital library technology, huge amount of rubbing from a stone inscription documents are scanned into the computer and stored as digital images, and have been made available to the general public through specialized web portals. Here are many Chinese rubbings sets often used by researchers. Such as, the East Asian Library׳s collection of Chinese rubbings in Berkeley University is second in number, outside of East Asia, only to that of the Field Museum of Natural History in Chicago. The Fine Arts Library currently houses 2602 individual East Asian rubbings, the majority of which are from China. Compared with paper rubbings, digital rubbings are more convenient for people to study.

Because the stones are not smooth originally and these inscriptions might have suffered from natural erosion during for thousands of years׳ storage or serious destruction when they are excavated and moved, most of them are not integrated and the original characters have been covered by many maculas, let alone the damage caused by having been tamped in the process of taking thousands of rubbings. Owing to the characteristics of stone inscriptions and the acquisition of objective factors of the environment, the original inscription images are filled with so much image noise that it may seriously affects the observation and research for inscription images, as shown in the two examples of rubbed calligraphy documents in Fig. 1, Fig. 2. In addition, there are other kinds of difficulties appearing in these images as different font types and sizes in the words, underlined and/or crossed-out words, etc. The combination of all these problems contributes to make the recognition process become very difficult, and hence, the preprocessing module quite essential.

The degradation of Chinese calligraphy images aesthetically affects the human perception and concomitantly the processes of feature recognition, segmentation, edge detection, etc. For correct interpretation of these images, restoration techniques are employed. Image denoising is one of the important fields in the restoration arena. The purpose of denoising is to obtain a good estimate of the original image from its degraded version and at the same time to preserve complex structures of images such as edges and textures. It is a difficult task to undertake, because the noises are randomly distributed in size and shape, and denoising sometimes may destroy simultaneously the characteristic parts of strokes simultaneously, such as the stroke tips and corners.

In the paper, we formulate this difficult rubbing image deblurring problem as an image denoising problem, using a pair of really rubbing images. The six classic denoising algorithm have been researched and compared, which are Anisotropic Diffusion filter, Wiener filter, TV(Total Variation, TV) minimizations, NLM (Non-Local Means, NLM), Bilateral filtering and Wavelet denoising. Finally, really rubbing images validate image denoising and analyzed the performance of the various algorithms have been described.

The article is organized as follows. Section 2 briefly describes the theory basis that different denoising methods and the relative researches are summarized. Five assessment parameters for image denoising have been explained in Section 3. Section 4 gives the experiments results of five different denoising methods, and five performance measurement criteria, such as PSNR, MSE, SNR, UQI and SSIM, are discussed. Section 5 gives a simple conclusion.

Section snippets

Relative works

In the late 1980s, the prevalence of fast computers, large computer memory, and inexpensive scanners fostered an increasing interest in document image analysis. An increasing number of Chinese calligraphy document images of different qualities are being scanned and archived. After document input by digital scanning, pixel processing is first performed (also called preprocessing and low-level processing in other literature) [10]. This level of processing includes operations that are applied to

Assessment parameters

The human eye is the only one able to decide if the quality of the image has been improved by the denoising method. To quantify the performance of the proposed algorithm, various parameters are employed such as PSNR, MSE, SNR, UQI and SSIM [11]. All the measures concerned are briefly introduced below.

  • (1)

    PSNR

    PSNR (Peak Signal to Noise Ratio) is defined as the ratio of the variance of the noise-free signal to the mean squared error between the noise-free signal and the denoising signal [51]. PSNR is

Experimental results and discussions

To investigate the performance of the different denoising methods, the images which come from library of Harvard are chosen as input images with resolution of 410*454 pixels and 256 intensity levels. We now present numerical experimental results and comparisons to illustrate the valid performance of the different denoising algorithms. Fig. 4, Fig. 5, Fig. 6 show the original image and the output images processed after filtering processing.

For the Wiener filter using MATLAB׳s wiener2 function,

Conclusions

The text images obtained through rubbing were featured by many fuzzy details, bad effect and so on, so it might lose more details in the traditional handling process. Though there are many kinds of image denoising algorithm, we have compared our method to seven well-known and widely available denoising algorithms: Anisotropic Diffusion Filter, Wiener filter, TV (Total Variation), NLM (Non-Local Means, NLM), Bilateral filtering, hard VisuShrink thresholding, soft VisuShrink thresholding,

Acknowledgments

This work was supported by the National Natural Science Foundation of China (Grant no. 61472173), the Grants from the Science and Technology Planning Project of Jiangxi Province of China, No. 20111BBG70032-2.

Zhi-Kai Huang received the B.Sc. degree in Mechanical Manufacture and Automation from the Kunming Polytechnic University, Kunming, China, in 1990, the M.Sc. degree in Mechanical Engineering from Central Southern University, Changsha, China, in 2001. In 2006, he received Ph.D. degrees in Control Theory and Control Engineering from University Science and Technology of China (USTC). From 1990 to 1998, employed as an assistant to the Chief Engineer of the Tongling Nonferrous Metals Group Holdings

References (56)

  • ...
  • D. Huang et al.

    A neural root finder of polynomials based on root moments

    Neural Comput.

    (2004)
  • B.K. Karch et al.

    Adaptive Wiener filter super-resolution of color filter array images

    Opt. Express

    (2013)
  • D.S. Huang et al.

    Zeroing polynomials using modified constrained neural network approach

    IEEE Trans. Neural Netw.

    (2005)
  • H. Om et al.

    An improved image denoising method based on wavelet thresholding

    J. Signal Inf. Process.

    (2012)
  • S.H. Chan et al.

    An augmented Lagrangian method for total variation video restoration

    IEEE Trans. Image Process.

    (2011)
  • K. He, J. Sun, X. Tang, Guided image filtering, in: Proceedings of Computer Vision–ECCV 2010, Springer, Berlin,...
  • S.M. Kay

    Fundamentals of Statistical Signal Processing, Volume III: Practical Algorithm Development

    (2013)
  • T.F. Chan et al.

    The digital TV filter and nonlinear denoising

    IEEE Trans. Image Process.

    (2001)
  • X.F. Wang et al.

    A novel density-based clustering framework by using level set method

    IEEE Trans. Knowl. Data Eng.

    (2009)
  • L.P. Yaroslavsky

    Digital Picture Processing, an Introduction

    (1985)
  • A. Buades et al.

    A review of image denoising algorithms, with a new one

    Multiscale Model. Simul.

    (2005)
  • D. Huang

    Radial basis probabilistic neural networks: model and application

    Int. J. Pattern Recognit. Artif. Intell.

    (1999)
  • J. Zhang et al.

    Denoising of Chinese calligraphy tablet images based on run-length statistics and structure characteristic of character strokes

    J. Zhejiang Univ. Sci. A

    (2006)
  • D.S. Huang

    A constructive approach for finding arbitrary roots of polynomials by neural networks

    IEEE Trans. Neural Netw.

    (2004)
  • D.L. Donoho

    De-noising by soft-thresholding

    IEEE Trans. Inf. Theory

    (1995)
  • S. Grace Chang et al.

    Adaptive wavelet thresholding for image denoising and compression

    IEEE Trans. Image Process.

    (2000)
  • H.J. Yu et al.

    Normalized feature vectors: a novel alignment-free sequence comparison method based on the numbers of adjacent amino acids

    IEEE/ACM Trans. Comput. Biol. Bioinform.

    (2013)
  • Cited by (22)

    • Restoration of stained old manuscripts via a hybrid wavelet and bilateral filtering system

      2017, Journal of King Saud University - Computer and Information Sciences
      Citation Excerpt :

      It is also used in other areas such as mesh fairing (Fleishman et al., 2003), volumetric denoising (Wong et al., 2004), optical flow and motion estimation (Xiao et al., 2006), and video processing (Bennett and McMillan, 2005). Authors of Huang et al. (2016) made a quantitative comparison between five denoising algorithms for Chinese calligraphy images filtering. In paper (Hedjam and Cheriet, 2013), they used multispectral imaging system to restore historical documents for the purpose of their recognition.

    • OraclePoints: A Hybrid Neural Representation for Oracle Character

      2023, MM 2023 - Proceedings of the 31st ACM International Conference on Multimedia
    • A Comparative Study of Several Spatial Domain Image Denoising Algorithm

      2023, Lecture Notes on Data Engineering and Communications Technologies
    View all citing articles on Scopus

    Zhi-Kai Huang received the B.Sc. degree in Mechanical Manufacture and Automation from the Kunming Polytechnic University, Kunming, China, in 1990, the M.Sc. degree in Mechanical Engineering from Central Southern University, Changsha, China, in 2001. In 2006, he received Ph.D. degrees in Control Theory and Control Engineering from University Science and Technology of China (USTC). From 1990 to 1998, employed as an assistant to the Chief Engineer of the Tongling Nonferrous Metals Group Holdings Co.,Ltd, Anhui province, China. In 2007, he joined the Nanchang Institute of Technology (NIT). Currently, he is a Professor of College of Mechanical and Electrical Engineering in NIT, Nanchang, China. His research interests include pattern recognition, image processing and data mining.

    Zhi-Hong Li received the B.Sc. and M.Sc. degree in Computer and Science Technology from Huazhong University of Science and Technology, Wuhan, China, in 1984 and 2004, respectively. Currently, she is a Professor of College of Mechanical and Electrical Engineering in NIT, Nanchang, China. His research interests include pattern recognition, Fault diagnosis of hydraulic generator sets.

    Han Huang, he is currently studying the B.Sc. degree in Industrial Engineering at the Harbin Institute of Technology, Harbin, China. His research interests include industrial engineering, image processing.

    Zhi-Biao Li received the B.Sc. degree in Mechanical Manufacture and Automation from Changchun University of Technology, Changchun, China, in 2003 and the M.Sc. degree in Mechanical Engineering from Nanchang University, Nanchang, China, in 2014. He started to teach in Nanchang Institute of Technology in 2003 and is currently a lecturer here.

    Lin-Ying Hou received the B.Sc. degree in Computer and Science Technology from Nanchang Institute of Technology, Nanchang, China, in 2004. She joined the Computer Science & Technology Department of Nanchang Institute of Technology as a lecturer in 2007. Her research interests include image analysis, visual and geometric computing.

    View full text