Evaluation of Current Documents Image Denoising Techniques: A Comparative Study

Abstract

The research to come out with effective and noise-free images is still fresh despite several years’ efforts. Even though, recently, several methods have been proposed, each approach has its own merits and limitations. Moreover, outstanding performances are exhibited for the tailored applications but fail in general and create several flaws and blur images of fine structures. The purpose of this study is to compare achievements thus far in the area of nonnatural documents image denoising approaches. Accordingly, existing techniques are organized into nonhomogeneous categories and detailed comparison is exhibited with examples. Finally, remaining problems and possible future directions in the domain of document images denoising are suggested.

INTRODUCTION

Noise removal from natural and nonnatural images is a chronological as well as a persistant problem in the domain of image processing. It cannot be avoided, and noised images reduce optimal character recognition (OCR) accuracy significantly. Hence, it is normally included in all preprocessing steps. These applications include OCR, offline script recognition, character segmentation, object tracking, image segmentation, and classification. Paper documents are usually scanned and these scanned images are tainted with noise. The noise usually comes from different sources such as dirt on paper or lens, moisture on the lens, or physical dealing of papers. Moreover, noise is also introduced by transmission errors and compression. This noise might cause serious problems in the transferring of these images into ASCII codes during the OCR process. Additionally, significant features of the images are corrupted because of noise and causes for image degradation that significantly reduce OCR accuracy. Hence, efficient image denoising techniques are mandatory in order to compensate for data corruption to level out further processing of image segmentation and classification.

Document images might be infected by several classes of document noise because of physical dealing of papers or low quality of the scanning devices (Rehman, Saba, and Sulong 2010). These noises consist of salt and pepper noise (Abdel-Dayem, Hamou, and El-Sakka 2004), shadow noise (Ping, Lihui, Alex 2000), salient noise, and white-line dropout noise (Huang and Zhu 2010).

CURRENT IMAGE DENOISING APPROACHES

The literature is full of image denoising approaches and it is difficult to organize them into heterogeneous classes. Currently, all techniques reported as state of the art could be organized into two classes: spatial filtering methods and transform domain filtering methods (Motwani, Gadiya, and Motwani 2004), as exhibited in Figure 1.

FIGURE 1 Hierarchical structure of current image denoising approaches (Motwani et al. 2004).

Spatial Filters

The spatial filters are composed of linear and nonlinear filters, detailed in the following sections.

Linear Filters

The linear filters are easy to implement, however, they are best applicable for random noise detection and removal. Other than that, they blur edges and remove fine detail and lines, so the efficiency is not good (Liying, Lixin, and Chew 2001). To rectify inherited drawbacks of mean filters, the Wiener filtering technique is implemented (Farahanirad et al. 2011). Nonetheless, the Wiener filtering technique removes the noise, yet for text and blurring of edges, it can work only under the condition of smooth underlying signals. Furthermore, it cannot work without details of the noise spectra and the actual image (Xing-mei et al. 2007; Saba, Rehman, and Elarbi-Boudihir 2011). Figure 2 exhibits results of linear filters on coin images.

FIGURE 2 Linear filters’ effects.

Nonlinear Filters

Nonlinear filters are proposed to eliminate the shortcoming of linear filters to preserve structural information of an image and are used to denoise images without identifying noise. The most popular nonlinear filters are morphology and median filters.

The morphological operations are of two types: erosion and dilation, employed for salt-and-pepper noise identification and removal (Chen et al. 2011). However, morphological operations are suitable for natural images, yet, at the same time, create several problems for nonnatural images. Consequently, numerous researches have been conducted to enhance morphological techniques (Zheng and Kanungo 2001; Wang and Wu 2009; Mehrotra et al. 2012). Although significant efforts have been made, morphological techniques still have serious drawbacks (Al-Khaffaf, Talib, and Salam 2008; Bovik 1987).

Therefore, they are not feasible for nonnatural images (He et al. 2008; Nobi and Yousuf 2011), vector-median-filtering (Barni et al. 2009; Rehman, Kurniawan, and Saba 2011), image continuation algorithms, rank order filtering (Abreu et al. 1996), relaxed median filter (Hamza et al. 1999), switching median filter (Saba and Rehman 2012), progressive switching median filter (Wang and Zhang 1999), and improved switch median filter (Zhang and Karim 2002; Nikolova 2004; Chan, Ho, and Nikolova 2005; Dong and Fang 2006; Dalong, Simske, and Mersereau 2007). Figure 3 exhibits the effects of nonlinear filters on a document image.

FIGURE 3 Effects of nonlinear filters: (a) original image (b) filtered image.

Transform Domain Filtering

Transform domain filtering methods are categorized into nonadaptive and data-adaptive transform filters, based on their functionality. The first category of transform domain filtering is well discussed and experimented in the literature of the state of the art. Additionally, it is further subdivided into spatial-frequency filtering and wavelet domain.

Spatial-Frequency Filtering

Spatial-frequency filtering approaches employ low-pass filters and fast Fourier transform (FFT). Hence, such techniques are sluggish and highly dependable on the cut-off frequency and performance of filter function. Moreover, fast Fourier transform-based techniques generate artificial frequencies (Buades, Coll, and Morel 2005), as exhibited in Figure 4.

FIGURE 4 Effects of spatial-frequency filtering: (a) original image (b) filtered image.

Wavelet Domain Filters

Wavelets are popular because of their efficient image denoising output (Bala and Ertuzun 2002). Hence, wavelet transform-based approaches are effective for natural image denoising. Accordingly, the best results are yielded in the literature (Kuo and Johnston 2002; Choi and Baraniuk 2004; Zhou, Zhou, and Stewart 2006). However, these approaches failed to produce better results for nonnatural documents images (Dalong 2009), as exhibited in Figure 5.

FIGURE 5 Effects of wavelet domain filters: (a) original image (b) filtered image.

The authors recommend in-depth reviews in Ahmed Mashaly et al. (2010), Seo et al. (2007), and Saba, Sulong, and Rehman (2011).

Data-Adaptive Transforms

This approach has proven to be a better option for natural and nonnatural image denoising applications (Mundher et al. 2014). Effects of data-adaptive transforms on natural and nonnatural images are exhibited in Figure 6. However, these filters are also slow and high-memory demanding (Mahmoudi and Sapiro 2005). Recently, image denoising techniques reported in the literature with regard to the state of the art are employing support vector regression (Dalong 2009) and yield promising results for both categories (natural and nonnatural images).

FIGURE 6 Effects of data-adaptive transforms: (a) original image (b) filtered image.

CONCLUSION AND RECOMMENDATIONS

Efficiency of image denoising techniques could be measured using peak signal-to-noise ratio (PSNR) and signal-to-noise ratio (SNR). Additionally, visual impacts of the enhanced images are also used as a yard stick in a few of the image denoising research activities. Conversely, existing approaches consider the noise model to be Gaussian, although, in fact, this supposition is not always valid because of the huge variety of natural images, nonnatural images, and the source of the noise. Moreover, a perfect image denoising algorithm needs a priori knowledge of the noise in order to gather information about its variance to develop a model. This noise model is employed to compare the performance with different algorithms. However, practically, it is not possible to have an original image free of noise.

Certainly, wavelet-transform-based approaches evolved with effective tools to denoise an image because of their valuable attributes such as sparsity, multiresolution, and multiscale nature. Hence, potential achievements are reported in the literature based on the aforementioned approaches (Rehman and Saba 2011a,b).

Currently, multiwavelets filters improve only image denoising accuracy; yet, they increase memory consumption and are demanding with regard to high processor times. Additionally, computational complexity has been ignored in previous researches and researchers improved only accuracy, so this parameter should also be a focus of future research.

Additionally, the choice of primary resolution while using wavelet transform is a critical issue, and it might affect the success of the application (Rehman and Saba 2014). Therefore, researchers must be careful in selecting a true image denoising approach for true application and comparison on benchmark databases (Rehman et al. 2009).

Future research should focus on fused approaches; the mixture with intelligent approaches might produce promising results, particularly for nonnatural images. Actually, there is an utmost need to propose image denoising approaches that are equally suitable for both natural and nonnatural images. Currently, image denoising approaches are lacking these features and most are customized so one technique is suitable for its particular environment and fails for different images (Saba et al. 2014).

FUNDING

The authors extend their appreciation to the Deanship of Scientific Research at King Saud University for funding this work through research group no. RGP-264.