Вы находитесь на странице: 1из 7

Automated System for Denoising Gray-Scale Images Using Image Priors

Stanley Jebakumar Dept. of ECE, Karunya University Narain Ponraj Asst. Prof/ECE Karunya University Samuel Manoharan Associate Prof/ECE BCET Merlin Mercy Asst. Prof/CSE Krishna College of Engg.

Abstract The first part of this paper proposes an Image Priors technique for image denoising via field of experts model. This is a new technique for image denoising and different from other wavelet or transform based techniques. Image priors can be learned from a set of natural images or sometimes from a single image or a collection of similar images. The basic model and the higher order MRF priors for colour images are also in the first part of the paper. The second part of this paper discusses about the methodology followed for denoising and the implementation of the technique. The results of the experiments are shown and they are discussed with their recorded signal to noise ratio values. While the model is trained on a generic image database and is not tuned toward a specific application, the results obtained here compete with and even outperform some specialized techniques.

1. Introduction
Digital images play an important role both in daily life applications such as satellite television, magnetic resonance imaging, computed tomography as well as in areas of research and technology such as geographical information systems and astronomy. Data sets collected by image sensors are generally contaminated by noise. Imperfect instruments, problems with the data acquisition process, and interfering natural phenomena can all

degrade the data of interest. Furthermore, noise can be introduced by transmission errors and compression. Thus, denoising is often a necessary and the first step to be taken before the images data is analyzed. It is necessary to apply an efficient denoising technique to compensate for such data corruption[1]. Image denoising still remains a challenge for researchers because noise removal introduces artifacts and causes blurring of the images. In this work, we describe different methodologies for noise reduction (or denoising) giving an insight as to which algorithm should be used to find the most reliable estimate of the original image data given its degraded version. Noise modelling in images is greatly affected by capturing instruments, data transmission media, image quantization and discrete sources of radiation. Different algorithms are used depending on the noise model. Most of the natural images are assumed to have additive random noise which is modelled as a Gaussian. Speckle noise[2] is observed in ultrasound images whereas Rician noise[3] affects MRI images. The scope of this paper is to focus on noise removal technique for natural images using image priors.

2. Image Priors
Image priors can be learned from a set of natural images or sometimes from a single image or a collection of similar images. These priors can then be used to perform various inference tasks, such as

image denoising, image segmentation, motion estimation, object removal and super-resolution. In this work we study one of the two image prior methods given below and compare their results with other techniques. The two methods are: 1) Fields of Experts (based on Marko Random Field) and 2) Epitome (a miniature version of an image or a set of images that describe an image) In the first approach (Fields of Experts) the authors develop a method for learning rich Markov random field (MRF) image priors by exploiting ideas from sparse image coding. The resulting Field of Experts (FoE) models the prior probability of an image in terms of a random field with overlapping cliques, whose potentials are represented as a Product of Experts[4]. They show how the model is trained on a standard database of natural images[5] and develop a diffusionlike scheme that exploits the prior for approximate Bayesian inference. This model can be used for image denoising, image inpainting[6] and Optical Flow[7]. Despite the generic nature of the prior and the simplicity of the approximate inference, the results obtained by the model are state of the art and a significant step up from the earlier MRF (Markov Random Field) approaches.

The FoE model works well for the monochromatic images but can the same be extended to colour images where the dimensionality is high. To explore this possibility we also look into a novel approach suggested by McAuley[8] which deals with the high dimensionality of data using FoE for the colour images. We also conduct experiments to learn filters and alphas using their suggested method and perform inference with the focus on denoising to compare results with the Epitome method. Additionally, we extend the FoE colour image model to do video denoising. The authors of the Epitome model have presented epitome as a suitable model for various imaging applications. One such application is denoising and authors have presented experimental work for the same. However there is little statistical data available that can be used to show the goodness of the model for denoising in general and to compare results with other models, for example, with FoE or other state of the art methods. In this work we use epitome with the focus on denoising and get statistical results to compare with the FoE model. While implementing the FoE model we also analyse the advantages and limitations of the model.

4. Field of Experts Model


The statistics of small image patches have received extensive treatment in the literature. The Field of Experts model proposed by Roth and Black[6] introduced a generic Markov random field image prior over extended neighbourhoods which is capable of representing a richer prior structure for images. In order to cope with the large clique sizes involved, they parameterize the potential functions in

3. Image Denoising
Both, Field-of-Experts and Epitome models, are based on image priors. The paper is based on creating an image denoising system that would exploit and implement these two models and while doing so we would compare the two approaches and the results obtained.

terms of filters and associated coefficients. Learning in the high dimensional space of the clique then becomes simply learning the filters and the coefficients. Before discussing the Field of Experts model further, let us take a look at the Product of Experts model by Welling [9]. Using PoE framework it is possible to model high dimensional probability distributions by taking the product of several expert distributions, where each expert works on a low-dimensional subspace which is relatively easy to model. Usually, experts are defined on linear onedimensional subspaces (corresponding to the basis vectors in sparse coding models). It is worth noticing that projecting an image patch onto a linear component is equivalent to filtering the patch with a linear filter described by Ji. Based on the observation that responses of linear filters applied to natural images typically resemble a Student t-distribution, Welling et al. propose the use of Student-t experts[9]. The full Product of tdistribution (PoT) model can be written as

with One important property of this model is that both the i and the image filters Ji can be automatically learned from training data.

5. Basic Model
The PoT model described in the preceding section provides an elegant and powerful way of learning prior distributions on small image patches. However, the results do not generalize immediately to give a prior model for the whole image. Also, simply making the patches bigger is not a viable solution because of the following reasons: (1) The number of parameters to learn would be too large; (2) The model would only work for one specific image size and would not generalize to other image sizes; (3) The model would not be translation invariant, which is a desirable property for generic image priors. To understand this, let the pixels in an image be represented by nodes V in a graph G=(V,E), where E are the edges connecting nodes. Defining a neighbourhood system that connects all nodes in an m m rectangular region, every such neighborhood centered on a node (pixel) k=1, . . . , K defines a maximal clique x(k) in the graph. The Hammersley-Clifford theorem establishes that the probability density of this graphical model can be written as a Gibbs distribution

where i = {i,Ji} and the experts the form,

have

and Z() is the normalizing, or partition, function. The i are assumed to be positive, which is needed to make the proper distributions, but note that the experts themselves are not assumed to be normalized. It will later be convenient to rewrite the probability density in Gibbs form as

where x is an image and Vk(x(k)) is the potential function for clique x(k). It is assumed that the MRF is homogeneous; i. e., the potential function is the same for all cliques (or in other terms Vk(x(k))=V(x(k))). This property gives rise to translation invariance of an MRF model. Additionally, without any loss of generality it can be assumed that the maximal cliques in the MRF are square pixel patches of a fixed size; however, other non-square neighbourhoods could also be used[10].

Figure 1. Selection of the 5 5 filters obtained by training the Products-ofExperts model on a generic image database.

6. Higher order MRF priors of Color Images


McAuleys results are built on Roths results by generalizing the Fields of Experts model to a prior model for colour images. This means that tasks like image denoising, inpainting or super-resolution do not need to be done independently for each channel, but can be performed directly over a multiband image in such a way that the correlations between the different channels can be exploited by the model. Although in their model they use a naive learning procedure, they present experimental results that show improvements of the colour image prior over the state-of-the-art monochromatic priors reported by Roth and Black[6]. It is possible to expect these improvements to become more significant if their learning algorithm can be optimized further. The FoE model is a high-order Markov random field model where the cliques are square image patches of typically 3x3 and 5x5 pixels. Figure 1 shows a selection of the 24 filters obtained by training this PoE model on 5 5 image patches whereas the Figure 2 shows for the FoE model.

Figure 2. Selection of the 5 5 filters obtained by training the Fields-ofExperts model on a generic image database. The potential functions over these cliques are then assumed to be products of experts[4], i.e. products of individual functions f of the response of a filter Jf to the image patch xc:

These are assumed to be stationary (c:=), i.e. every clique in the image will have the same parameter vector ={Jf, f : 1 f F}. The particular form they postulate for the expert is related to the Student-T distribution, and reads

By invoking the Hammersley Clifford theorem, which states that the joint probability distribution of a Markov random field with clique set C can be written as

the model becomes

Both the Jfs and fs are learned as described in the next section. Once the model is learned, inference is performed by gradient ascent on the log-posterior.

7. Methodology
The methodology for denoising using FoE model can be explained with the help of the following block diagram:

taken as given for the second. This is done to get better results with good PSNR values. The denoising status can be known by printing the results after particular iterations using a if loop in the program. The gradient of the denoised image is calculated at the end of every iteration. When the denoising process is over, the resultant image is converted back to the RGB colour scale.

8. Results
The experiments conducted here assume a known noise distribution. We used an FoE prior with 24 filters of 5 5 pixels. We chose the update rate to be between 0.02 and 1 depending only on the amount of noise added, and performed 3000 iterations. While potentially speeding up convergence, large update rates may result in numerical instabilities, which experimentally disappear for 0.02. However, that running with large step sizes and subsequently cleaning up the image with 250 iterations with = 0.02 shows no worse results than performing the denoising only with = 0.02. Experimentally, the best results are obtained with an additional weight for the likelihood term, which furthermore depends on the amount of noise added. We automatically learn the optimal value for each noise level using the same training data set that was used to train the FoE model. This is done by choosing the best value from a small candidate set of s. The results obtained for the peppers image is shown in the figures below:

Figure 3. Block Diagram showing the methodology The FoE model to be used for the denoising procedure is selected first i.e. either 3x8 or 5x24. The image used for the experiment is read from the file location. The additive white Gaussian noise is added with the image according to value given. This process is carried out by the noise adder. Now, the denoising process starts initially by finding the - pair. In case the input image is in RGB colour scale, then for ease processing the image is converted to YCbCr colour space. Then the denoising process starts and continues till it completes the given number of iterations. Basically, in this method the denoised image of the first iteration is

Figure 4. Final Denoised output for the peppers image The PSNR values of the denoised images after every 250 iterations are shown in the figure below:

Figure 5.Denoised images with their PSNR values In the same way, the PSNR values in dB obtained for other standard grey scale images are tabulated in the following table:

1 2 5 10 15 20 25 50 75 100

Lena
47.84 42.92 38.12 35.04 33.27 31.92 30.82 26.49 24.13 21.87

Barbara
47.86 42.92 37.19 32.83 30.22 28.32 27.04 23.15 21.36 19.77

Boat
47.69 42.28 36.27 33.05 31.22 29.85 28.72 24.53 22.48 20.80

House
48.32 44.01 38.23 35.06 33.48 32.17 31.11 26.74 24.13 21.66

Peppers
47.81 42.96 37.63 34.28 32.03 30.58 29.20 24.52 21.68 19.60

9. Conclusion
The resulting Fields-of-Experts model is based on a rich set of learned filters, and is trained on a generic image database using contrastive divergence. We have demonstrated the usefulness of the FoE model with applications to denoising. Further this model can also be used for inpainting. The denoising algorithm is straightforward (approximately 20 lines of MATLAB code), yet achieves performance close to the best specialpurpose wavelet-based denoising algorithms. The advantage over the wavelet based methods lies in the generality of the prior and its applicability across different vision problems. We believe the results here represent an important step forward for the utility of MRF models and will be widely applicable. The results here are also applicable to image super-resolution, image sharpening, and graphics applications such as image based rendering[11] and others. There are many avenues along which the FoE model itself can be studied in more detail, such as how the size of the cliques as well as the number of filters influence the quality of the prior. Furthermore, it would be interesting to explore an FoE model using fixed filters (e.g. standard derivative filters or even random filters) in which only the expert parameters i are learned from data.

References
[1] Mukesh C. Motwani, Mukesh C. Gadiya, Rakhi C. Motwani & Frederick C. Harris, Jr. Survey of Image Denoising Techniques. [2] H. Guo, J. E. Odegard, M. Lang, R. A. Gopinath, I. W. Selesnick, and C. S.

Burrus, "Wavelet based speckle reduction with application to SAR based ATD/R," First Int'l Conf. on Image Processing, vol. 1, pp. 75-79, Nov. 1994. [3] Robert D. Nowak, Wavelet Based Rician Noise Removal, IEEE Transactions on Image Processing, vol. 8, no. 10, pp.1408, October 1999. [4] G. Hinton. Product of experts. ICANN, v. 1, pp. 16, 1999. [5] D. Martin, C. Fowlkes, D. Tal and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. ICCV, v.2, pp. 416423, 2001. [6] S. Roth & M. J. Black. Fields of experts: A framework for learning image priors. IEEE Conference on Computer Vision and Pattern Recognition, pp. 860 867, 2005. [7] S. Roth & M. J. Black. On the spatial statistics of optical flow. IEEE International Conference on Computer Vision, pp. 42 49,2005. [8] McAuley, J., T. Caetano, A. Smola and M. O. Franz. Learning High-Order MRF Priors of Colour Images. Proc. 23rd Intl. Conf. Machine Learning, 2006. [9] M. Welling, G. Hinton, and S. Osindero. Learning sparse topographic representations with products of Student-t distributions. NIPS 15, pp. 13591366, 2003. [10] D. Geman and G. Reynolds. Constrained restoration and the recovery of discontinuities. PAMI, 14(3):367383, 1992. [11] A. Fitzgibbon, Y. Wexler, and A. Zisserman. Image-based rendering using image-based priors. ICCV, v. 2, pp. 1176 1183, 2003.

Вам также может понравиться