A Simple and Fast Algorithm for Image Denoising Using the DCT

A. Khelalef, 2Y. Himeur, 1N. Benoudjit 1 Laboratoire dElectronique Avance, Universit de Batna, Batna, Algeria Email: khelalef.aziz@gmail.com, nbenoudjit@gmail.com 2 Department of Electronics University of Jijel Jijel, Algeria Email: Him.yassine@gmail.com
Abstract. In this paper we present a new simple and fast algorithm for image denoising using two-dimensional discrete cosine transform. The proposed method introduces a new threshold selection to shrink the DCT coefficients and use the cycle spinning algorithm in order to improve the denoising performances. Simulation results show that the proposed scheme is fast , simple to implement and outperform the different methods cited in the literature by giving better results for all benchmark images for different values of standard deviation of the Gaussian noise, those proprieties make the proposed technique suitable for real time applications. Keywords: image denoising, thresholding, DCT, cycle spinning algorithm.
1

1 Introduction
In many applications in computer vision and image processing, images are corrupted by Gaussian noise [1]; for this, the purpose of image denoising is how to extract the true image informations from the noisy data; i.e. found a compromise between noise reduction and preservation of the important image features. In this context many methods were proposed. Starting from the famous work of D.L Donoho [2], image denoising scheme based on transform techniques like DWT and DCT became a useful tool to denoise an image. Earlier, image noise reduction was performing by using filters, generally the denoising images suffer from blurring and edges degradation [3], recently and to resolve filters inability new techniques based on the discrete wavelet transform (DWT) where proposed. Donoho and al. [2] proposed a wavelet image thresholding for image denoising, where the principal is simple: i) performing the DWT decomposition in different levels, ii) performing a thresholding operation on the details and iii) applying an inverse DWT to reconstruct the denoised image. Nowadays, and to outperform the conventional DWT techniques, neural networks are used to optimize the thresholds. Zhang in [4] and [5] proposed a space scale adaptive technique called Thresholding Neural Network (TNN) to reduce the Gaussian noise.The TNN is used to obtain the optimized thresholds for shrinking the DWT (Discrete Wavelet Transform) coefficients stream. In [6] a new TNN subband adaptive image denoising was proposed, where the author propose to use separate TNN for each subband. Unfortunately those kinds of methods suffer from threshold selection, algorithms complexity and heavy program execution, so they are not suitable for real time applications. In this paper, we propose a new simple and fast image denoising method based on the DCT. The technique is based on using the two-dimensional (2D) discrete cosine transform DCT to obtain the elementary frequency components, then applying a hard thresholding to the DCT coefficients using the appropriate proposed threshold. The thresholding process in combined with a cycle spinning algorithm to improve the denoising results [7]. The proposed method shows strong proprieties, the algorithm is simple, easy to implement, fast and suitable for real time applications. This paper is organized as follows. In section two, we give an introduction to discrete cosine transform and in the third section, we present brief introduction into image denosing in the transform domain. In the fourth section the cycle spinning algorithm is described, the proposed method will be given in section five. Simulation results and discussion are given in the six sections; finally, section seven contains concluding remarks and perspectives.

2 Discrete cosine transform

Like other transforms, the Discrete Cosine Transform (DCT) attempts to decorrelate the image data. After decorrelation each transform coefficient can be encoded independently without losing compression efficiency. This section describes the DCT and some of its important properties. The most common DCT definition of a 1-D sequence of length N is [8]:

for

. Similarly, the inverse transformation is defined as

for

. In both equations (1) and (2)

is defined as

Thus, the first transform coefficient is the average value of the sample sequence. In literature, this value is referred to as the DC Coefficient. All other transform coefficients are called the AC Coefficients. The objective of this study is image denoising using 2D DCT; this necessitates the extension of ideas presented previously to a two-dimensional space. The 2-D DCT is a direct extension of the 1-D case and is given by [8]:

For

and

and

are defined in (3), the inverse transform is defined as

For An example of the DCT on cameraman image is presented in figure (1).

Fig.1. exemple of DCT aplied on cameraman image: from left to right, originale, DCT, DCT using windows of 4x4, DCT using windows 8x8.

3 Image Denoising in the Transform Domain

In the case of a Gaussian noise, the problem is how to recover a function from noisy data And The application of a transform like the DCT on the function g gives: To recover data from , all coefficients are threshold depending on the noise contribution. Since the famous thresholds function (soft and hard) of Donoho [2], a lot of thresholding function were proposed in the denoising image field. [9]:

4 Cycle Spinning Algorithm

We use the cycle spinning algorithm proposed by Coifman and Donoho [7].Since the first thresholding algorithm proposed by Donoho [2] exhibits visual artifacts and oscillations in the vicinity of signal discontinuities, called pseudo-Gibbs phenomena. This method utilizes the shift variant property of the transform. In this algorithm by using different shifts of the noisy image, we can compute different estimates of the unknown (image), and then average these estimates [7]. If we denote the 2-D circular shift by , the DCT transform by , and the threshold operator by , the cycle spinning will be performed as:

where [7].

and

are the maximum number of shifts which would cause an improvement in the denoising process

5 Proposed Method
The proposed method is very simple to implement, at this point we present the algorithm. The thresholding process is embedded in the cycle spinning algorithm, we fix the maximum number of the horizontal and the vertical shift by 4, the algorithm is presented below: For =1 to 4 For =1 to 4 1. 2. 3. 4. 5. 6. 7. End End 8. Average the saved images to obtain the free noise image Shift the noisy image horizontally by and vertically by . Apply the windowing DCT to the shifted noisy image using 4 by 4 masks. Calculate the threshold value using the equation 9 given by : Threshold the DCT coefficients using a hard thresholding function given by equation 10 (figure 2). Apply the inverse (IDCT). Shift inverse the free noise image version horizontally by Save the new free image version. and vertically by .

We propose to use a hard thresholding function proposed by D.L Donoho in [2], given by equation 10 and represented in figure 2 below:

Where;

are the noisy DCT coefficients, is the threshold and

is the thresholding function.

Fig.2. hard thresholding.

6 Experimental Results
The proposed algorithm has been tested on different gray scale images (of size 512 512 and 256 256, with 8 bpp). These images were corrupted with different standard deviation of Gaussian noise (=10, 15, 20, 25 and 30). To evaluate the performance of the proposed method, the results are compared to different wavelet denoising schemes. In this paper, we used the standard evaluation criterion in Gaussian noise methods that is Peak Signal to Noise Ratio (PSNR) and is defined by [9]:

: Original image. : Denoised image. : Image dimension.

Image

10 15 20 25 30 10 15 20 25 30 10 15 20 25 30 10 15 20 25 30

Noisy
28.1538 24.6113 22.0965 20.1931 18.5878 28.1291 24.6076 22.1232 20.1652 18.5792 28.1278 24.6364 22.1247 20.1715 18.5729 28.1449 24.6078 22.1009 20.1036 18.5796

Wiener2
33.5508 31.1125 28.9920 27.2295 25.7290 29.7587 28.1547 26.6780 25.3770 24.2290 32.1973 30.2249 28.4101 26.7963 25.3963 30.5864 29.0860 27.5811 26.2126 24.9617

BayseShrink [10]
33.4404 31.6148 30.3188 29.3731 28.6018 31.5230 29.3361 27.7740 26.5559 25.6210 32.4387 30.2710 28.8722 27.8366 27.0236 31.1318 28.6142 27.0871 25.8961 25.0992

NormalShrink [4]
33.5473 31.6357 30.2804 29.1985 28.3901 31.8695 29.3547 27.6277 26.3765 25.4386 32.4087 30.2440 28.7758 27.7150 26.8873 31.0047 28.7330 27.2275 26.0556 25.1378

New TNN [6]

33.7452 31.8606 30.5670 29.5121 28.7208 32.0576 29.6381 27.9456 26.6340 25.7363 32.6597 30.4663 28.9925 27.9429 27.1660 31.4361 29.0092 27.3252 26.0821 25.2519

Proposed
34.9355 32.9683 31.4492 30.3051 29.2816 33.2612 30.5516 28.5901 27.1779 26.1434 34.2335 31.9540 30.3904 29.0598 28.0070 33.1722 30.4968 28.6848 27.4487 26.5078

Lena
(512x512)

Barbara
(512x512)

Boat
(512x512)

Cameraman (256x256)

Table 1. PSNR results for various denoising methods on different benchmark images.

36

34

Wiener NormalShrin BayseShrink NewTNN proposed

32

PSNR (dB)

30

28

26

10

15

20

25

30

sigma

Fig.3. Comparaison of desoising performance (PSNR) for the four methods and the proposed method on Lena image.

The results are compared with four methods given in literature Wiener, BayseShrink [10], NormalShrink [4], and New TNN [6]. By a global comparison, the experimental results (table 1 and figure 3) show better PSNR performance for the proposed method comparing to the other methods for all images test and different standard deviation. Figures (4 and 5), show that the proposed method gives better details and edges preservation for high noise levels for all images used in simulation.

Fig.4. Image denoising results on Part of Lena Image (from top to bottom): Noisy (=20) , Wiener2, BayseShrink [10], NormalShrink [4], New TNN [6],proposed method,original.

Fig.5. Some gray scale image denoising using the proposed method. From Left Noisy image (=20), from right proposed method.

7 Conclusion
In this paper, we have proposed a new DCT image denoising using a new threshold and cycle spinning algorithm. In this technique we use a windowing DCT, where the coefficients are shrink using a hard thresholding function, On the other hand, the cycle spinning algorithm is used to improve the denoising results. The algorithm outperforms the existing methods objectively (PSNR) and subjectively (visual quality),for all the images used is simulation and for all standard deviation of the Gaussian noise. The proposed method is fast; simple to be implement, and suitable for real time applications In future works, we will base on using adaptive thresholding function and optimum thresholds to improve the denoising results.

References
[1] Fengxia Yan, Lizhi Cheng, and Silong Peng, A New Interscale and Intrascale OrthonormalWavelet Thresholding for SURE-Based Image Denoising, ieee signal processing letters, VOL. 15, 2008, pp,139-142. [2] D. L. Donoho, De-noising by soft-thresholding, IEEE Transaction on Information Theory, Vol. 41. pp. 613-627, May 1995. [3] A. buades, B. coll, and J. M. morel, a review of image denoising algorithms, with a new one, multiscale model. simul. Vol. 4, No. 2, pp. 490530. [4] X.-P. Zhang, Space-scale adaptive noise reduction in images based on thresholding neural networks, in: Proceedings of IEEE International Con ference on Acoustics, Speech, and Signal Processing, 2001, pp. 18891892 [5] X.-P. Zhang, Thresholding Neural Network for Adaptive Noise Reduction, IEEE Transaction on Neural Networks, Vol. 12 (3) (may 2001), pp. 567584. [6] Mehdi Nasri, Hossein Nezamabadi-pour, Image denoising in the wavelet domain using a new adaptive thresholding function, Neurocomputing, 72, 2009, pp. 1012-1025. [7] R. R. Coifman and D. L. Donoho, "Translation invariant denoising", in Wavelets and Statistics, Springer Lecture Notes in Statistics 103, New York, Springer-Verlag, pp. 125-150, 1994. [8] R. M. Zhao, H. Lian, H. W. Pang, B.N. Hu, A Watermarking Algorithm by Modifying AC Coefficies in DCT Domain, International Symposium on Information Science and Engieering, pp. 159-162. IEEE, 2008. [9] Karunesh K. Gupta and Rajiv Gupta, Feature Adaptive Wavelet Shrinkage denoising, IEEE - ICSCN 2007, MIT Campus, Anna University, Chennai, India. Feb. 22-24, 2007. pp.81-85. [10] S. Grace Chang, Bin Yu, Martin Vetterli, Adaptive Wavelet Thresholding for Image Denoising and Compression, ieee transactions on image processing, vol. 9, NO. 9, September 2000, pp. 1532-1546.