International Journal of Computer Applications (0975 – 8887) Volume 114 – No. 16, March 2015 33 Adaptive Edge-Preserving Image Denoising using Arbitrarily Shaped Local Windows in Wavelet Domain Paras Jain Jaypee University of Engineering & Technology Raghogarh, Guna 473226, MP, India Vipin Tyagi Jaypee University of Engineering & Technology Raghogarh, Guna 473226, MP, India ABSTRACT Image denoising is a well explored topic in the field of image processing. A denoising algorithm is designed to suppress the noise while preserving as many image structures and details as possible. This paper presents a novel technique for edge- preserving image denoising using wavelet transforms. The multi-level decomposition of the noisy image is carried out to transform the data into the wavelet domain. An adaptive thresholding scheme which employs arbitrary shaped local windows and is based on edge strength is used to effectively reduce noise while preserving significant features of the original image. The experimental results, compared to other approaches, prove that the proposed method is suitable for various image types corrupted by Gaussian noise. Keywords Wavelet transform; arbitrary shaped window; region-based approach; noise reduction; edge-preservation. 1. INTRODUCTION Image denoising is the process of obtaining the original image by reducing unwanted noise from a corrupted image. Various techniques [1-3] for noise reduction have been introduced in last few decades, many of them are based on linear spatial domain filters. Most linear filtering techniques proposed so far need some prior knowledge about the noise and the image characteristics. Usually, such details are not available and may be hard to estimate from the input noisy image. Linear filters are easy to implement and usually smooth the data to eliminate the noise; however, this process can produce blurring in images [3]. To handle the problems of linear filters, many non-linear edge-preserving methods have been developed and research is still continued in this direction [4]. These non-linear edge- preserving filtering methods can remove the noise more effectively while preserving the important image features such as edges. Wavelet-based non-linear methods have shown excellent potential in providing efficient edge-preserving image denoising since they provide a suitable basis for separating noisy signal from the image signal. The common approach for noise reduction in wavelet domain in to determine the multiscale wavelet decomposition of the noisy image and to modify the wavelet coefficients, thus obtained. Coefficients that are supposed to be noisy are modified by means of thresholding. Reconstruction from these modified coefficients then gives the desired denoised image. Numerous denoising methods follow such procedure of wavelet thresholding [5-14]. A major challenge in the wavelet shrinkage process is to estimate an appropriate threshold. Some well-known threshold estimation criteria are VisuShrink (non-adaptive) [6], SureShrink [7] (adaptive), BayesShrink [5, 17] (adaptive) and Cross Validation [18, 19] (adaptive). Once an appropriate threshold is estimated, the wavelet coefficients can be thresholded according to a shrinkage rule. A shrinkage (thresholding) rule at stage 3 defines the applicability of a threshold. The ultimate goal of a shrinkage rule is to preserve large coefficients which represent important signal features, while small coefficients can be thresholded without affecting the significant image features. Often used shrinkage rules are „keep-or-kill‟ hard thresholding and „shrink-or-kill‟ soft thresholding [16]. Most of the wavelet-based denoising methods often require the knowledge about the variances of signal and/or noise. Since these variances are usually unknown, assumptions are frequently made. A common solution to the problem is to estimate the variance from the data. Such variance estimation is popular for various denoising methods and achieves efficient denoising results with a low complexity. Numerous denoising approaches [10-13, 20] have taken benefit of joint statistical relationships among the wavelet coefficients by estimating the variance of a coefficient adaptively from a local neighborhood window consisting of coefficients within a subband. The size of the locally adaptive window also plays an important role in estimating the signal variance. In this paper, we present a new technique for noise reduction using wavelet transforms. A new locally adaptive thresholding scheme which involves estimation of thresholding parameters in arbitrarily shaped local neighborhood windows and is based on edge strength, is used to effectively suppress Gaussian noise while preserving relevant features of the original image. Experimental results show that the proposed method, when compared to well-known state-of-the-art denoising methods, is more suitable for various classes of images corrupted by Gaussian noise. The motivation behind using the concept of arbitrarily shaped (varying sized) local windows has come from the denoising approach suggested by Eom and Kim [11]. In [11], authors have been established that a locally adaptive window of nearly arbitrarily shape (i.e. varying size) can be more efficient in removing white Gaussian noise in comparison of fixed size local windows. The reason being with nearly arbitrarily shaped window, one can obtain more accurate local statistics of images. However, they had pointed that denoising results obtained with their approach were not so good in higher noise environment. Through the proposed approach, we have got better denoising results with effective edge- preservation. The remainder of this paper is organized as follows. Section 2 gives a brief review of related techniques; Section 3 contains the proposed method in detail; Section 4 gives experimental results, including a comparison with other denoising methods; Finally Section 5 summarizes the conclusions. 2. RELATED WORK A number of denoising approaches in wavelet domain have used the mechanism of locally adaptive windows. In this mechanism one can estimate the variance of each wavelet