An image representation is worth more than ten thousand words. A digital image is an image with spatial co-ordinates and amplitude values are finite in discrete quantities. The field of digital image processing refers to processing of two dimensional pictures by a digital computer.
Digital image processing has a broad spectrum of applications, such as robotics, biology, industry, astronomy, law enforcement, defense and automated inspection of industrial parts.
The area of applications of digital image processing are so varied that some form of organization is desirable in attempting to capture the breadth of this field. The simplest way to develop a basic understanding level of image processing applications is to classify the images according to their source.
1.1. Fundamental steps in Digital Image processing
The fundamental steps in digital image processing are acquisition, enhancement, restoration, compression, segmentation, representation and object recognition.
In Image acquisition, our objective is to generate digital images from sensed data. The output of most sensors is a continuous voltage waveform whose amplitude and spatial behavior are related to the physical phenomenon being sensed. To create a digital image we need to convert continuous sensed data into digital form.
Image enhancement is processing of manipulating an image so that the result is more suitable than the original for a particular application. Enhancement techniques are not suitable for all applications. For example, a technique that is helpful for improving X-ray images may not be the best technique for enhancing satellite images taken in the infrared band of the electromagnetic spectrum.
Fig. Fundamental Steps in Digital Image Processing
Image restoration is the process of removal or minimization of known degradations in an image. Random-noise reduction is carried out in the spatial domain using convolution mask and for modeling some important degradation, such as blur caused by motion during acquisition. A restoration filter that solves a given application in the frequency domain often is used as the basis for generating a digital filter that will be more suitable for routine operation using a hardware implementation.
Image compression deals with the techniques for reducing the amount of data required to represent an image. Image compression acts an important role in many areas. Because of the wide applications, data compression is of the great importance in digital image processing.
Image segmentation is a process that partitions an image into regions. Image segmentation is an essential preliminary step in most automatic pictorial pattern recognition and scene analysis application.
Noise in digital image
Noise removal plays an important role in digital image processing. The major source of noise in digital image occurs during image acquisition and transmission. The acquisition noise is generally additive white Gaussian noise
(AWGN) with very low variance and may be quite negligible for many engineering applications. It is mostly due to very high quality sensors. But in some applications such as remote sensing, biomedical instrumentation, etc., the acquisition noise may be high enough. This is due to the fact that the image acquisition system itself comprises of a transmission channel and such noise problems are measured as transmission noise. Then it may be concluded that acquisition noise is insignificant. Therefore, the researchers are mainly concerned with the noise in a transmission system. This noise appearance affects the original information of digital images. Image distortion is an important problem in image processing. Image distorted due to several kind of noise such as salt and pepper noise, Rayleigh noise, speckle noise and many other fundamental noise types in digital images. These noises may be arrived from a noise source established due to imperfection in the image capturing devices, weak focal length, scattering and other undesirable conditions may be appear in the atmosphere. This leads to in-depth study of noise and noise models for image denoising systems.
Image is an important source of information. But the noise in digital images creates undesirable effects such as blurred objects, distorts background views, shadows, unrealistic edges, corners and unseen lines. To reduce these effects, prior knowledge of noise models is necessary for further processing.
Types of noise models
Gaussian noise occurs in amplifiers or detectors. So it is also known as electronic noise. The probability density function of a Gaussian random variable z, is
Where, z represents the gray value, Âµ is the mean value of z, Ïƒ is its standard deviation.
The noise that occurs in coherent imaging of objects is known as speckle noise. Speckle noise is multiplicative noise, having granular pattern. Most of ultrasound images are corrupted by speckle noise.
It is expressed as:
g(m, n) = f(m, n) u(m, n) + Î·(m, n)
where, g(m, n) is corrupted image, u(m, n) is multiplicative component and Î·(m, n) is additive component.
The probability density function of Rayleigh noise is given by
The mean and variance of this density are given by
The Rayleigh density is useful in differentiating noise occurrence in range imaging.
Impulse valued noise
Impulse valued noise takes place in locations where a high transient (faulty switching) arises. The noisy pixels value corrupted by impulse noise are distributed in the range of [0, 255] for grayscale images. Impulse valued noise is also called as salt and pepper noise.
Though the image is not fully degraded by salt and pepper noise as choice of several pixel values are changed in the image. Although in noisy image, there is a possibilities of more than a few neighbors does not changed. Statistically the impulse valued noise drop the original data values. So data-drop-out is also a term used to refer to the impulse noise. The probability density function of impulse noise is given by
If b>a, intensity â€˜bâ€™ will shows a light dot in the image. Conversely, level of â€˜aâ€™ will appear like a dark dot. If either pa or pb is zero, the impulse noise is unipolar. Especially if the probabilities are equal and neither probability is zero, impulse noise values look like salt and pepper granules randomly spread over the image. Eliminating or reducing impulse noise is a very dynamic research area in image processing.
Fig. Probability Density Functions of Impulse Noise
Noise Models for Color Images
There are two types of noise in color images such as correlated color noise and uncorrelated color noise. In second type, noise affects the R, G and B planes separately with the given percentage. In first type, the occurrence of noise in a component of color pixel also depends on its occurrence in other components. It is caused in two steps. In the first step, noise is added in the similar way as uncorrelated noise. In the second step, for every noise-free component in any plane, it is checkered if other two parallel components in other planes are corrupt, and if so, noise-free component is made noisy based on the correlation factor.
Filtering is an important part of any signal processing system. Filtering consists of two types such as linear and non-linear filter. Linear filtering techniques are unsuccessful when the noise is non-additive and are not efficient in eliminating impulse noise. The non-linear filtering techniques are preferred for removing the impulse noise, because they are having the good edge and image detail conservation properties. Median filter is the most commonly used non-linear filter.
Denoising of images
Image denoising is an important issue in all image processing researches. The aim of image denoising is one of the essential task and pre-processing step in digital image processing. Usually, the significance of an image could be degraded by a lot of noise due to the undesired situations of image acquisition stage or during the transmission. The great challenge of image denoising is how to conserve the edges and all fine particulars of an image when reducing the noise.
Study of conventional filters
In Standard Median Filter, every pixel in the image is restored by the median value in its neighborhood as a result of which the enviable facts in the image are removed . SMF frequently shows blurring for huge window size and inadequate suppression for small window size.
The Weighted Median Filter  and the center weighted median filter  were proposed as remedy to progress the median filter by giving more weight to several elected pixels in the filtering window. Even though these filters can conserve more particulars than the median filter, they are still executed evenly across the image without consider whether the present pixel is noise free or not.
Directional Weighted Median Filter (DWM) [Dong and Shufang, 2007] is another method which eliminates random-valued impulse noise. This filter has established an impulse noise detection system that works on the differences of the middle pixel from its adjacent pixels along the four directions. The proposed method first identifies the impulse noise and then restores the noisy pixels. While eliminating noise, it employs the information of four directions in the window for detail conservation. It takes a 5x5 window and then regards the four major directions: horizontal, vertical and two diagonals, both having 5 pixels. Then it calculates the weighted variation in all directions and the least value is measured. The pixel is measured as noisy when this value is larger than a threshold. To eliminate the noise, standard deviation in all four directions is considered. Once the standard deviation has been established, the weighted median is calculated. It is accomplished by generous more weight on the direction of small standard deviation and noisy pixel is restoring by this median. The proposed method works fine in low to medium noise density but its presentation gets behind in case of high noise density.
An Adaptive Median Filter (AMF) is suitable for low and medium noise density levels . The above filter proposed that, the number of substitutions of corrupted pixel increases in higher noise densities and provides good noise removal performance for large window size. However the original pixel values and restored median pixel values are fewer correlated. As a result, the edges are smeared considerably. The Adaptive Centre Weighted Median Filter proposed in  is used to eliminate high density impulse noise and needs optimized threshold for both types of impulse noise.
The technique in Adaptive Center Weighted Median Filter (ACWMF) [Chen and Hong, 2001] is, adaptive operator makes approximates by developing the current pixelâ€™s difference from the outcome of center weighted median (CWM) [Ko and Yong, 1991] filters with different weights on the center. Switched method based on impulse detector is used for this purpose. Center weighted median filter is used since it has the capability to form a more common filtering operator. The proposed filter uses the difference among the output of CWM filters and the concerned pixel to identify impulse noise. Even though the proposed filter gives some better results in terms of noise suppression but destroying of fine pixels is more and its outcomes is generally a poor performance.
The Modified Decision Based Unsymmetric Trimmed Median Filter (MDBUTMF)  is proposed to remove high density salt and pepper noise in digital images. The above proposed filter assumes the maximum value (i.e. 255) and the minimum value (i.e. 0) in the dynamic range as salt noise and pepper noise respectively. The major problem is that at higher noise densities, if all the pixels in the 3 x 3 filtering window are corrupted either by salt noise value (i.e. 255) or by pepper noise value (i.e. 0), than it uses mean values of all the element in the filtering window to restore the noisy pixel which is also a noisy value i.e. 0 or 255.
Switching Median Filter [8-13] proposed a noise detector to establish that the middle pixel of a specified filtering window is corrupted or not. If the noise detector is identifying that the middle pixel is corrupted, then the outcome of the structure is switched to the output of the noise filter, which has the replaced value for the corrupted pixel. If the middle pixel is recognized as uncorrupted, which means that there is no requirement to perform filtering, the noise removal operator is avoided and the output of the structure is switched directly to the input. This approach has been engaged to extensively improve the routine of conventional median based filtering and a number of median based filters developing different impulse detection methods have been considered. Even though, progressive switching median filter  performs proficiently, it is time intense and computationally complex as an effect of which its hardware execution turns into difficult.
The Boundary Discriminative Noise Detection Filter (BDND) [Ng and Kai-Kuang, 2006] which is integrated into the framework of switching median filter has been proposed. The key success of the method is essentially due to highly precise noise detection capable by the BDND algorithm. Together with additional developments added from the post-detection filtering phase, the whole switching median filtering performance has yielded a better performance to that of the ideal-switching case constantly. The proposed noise detection system sometimes creates false alarms which corrupted the filtering performance. Also, the proposed method does not take concern of edges in the image which can cheat the noise removal algorithm.
Multiple Decision Based Switching Median (MDSM) filtering scheme has been observed to remove impulse noise by means of global and local statistics . Determining the accurate value of threshold for a given image is a difficult task in single threshold Switching Median Filtering. In addition, the single threshold value cannot be expected to give up optimal performance over the complete image since the images are non stationary processes. In order to avoid computational difficulty, multiple thresholds switching median filtering scheme (MTSMFS) has been examined . Obtainable switching-based median filters [17-27] are generally found to be non-adaptive to noise density deviations and prone to misclassifying pixel uniqueness. This exposes the significant need to develop a difficult switching scheme and median filter.
Artificial Intelligence Technique Approaches
Artificial Intelligence Technique approaches have a number of computational schemes in computer sciences, machine learning and huge engineering disciplines. Artificial intelligence technique tries to search, model, and analyze very complicated occurrence; those for which conventional methods doesnâ€™t give low cost and absolute solutions. Early computational methods were capable to exactly model and examine simple systems. Complex systems are now rising in design, digital imaging, engineering community, etc. which remain complicated for conventional analytical and mathematical techniques. A lot of the classical mathematical models have been very productive but difficult. A significant area of artificial intelligence technique contains Fuzzy logic and artificial neural networks (ANN).
Fuzzy logic may be defined as an addition of classical logical system [Yager and Lotfi, 1992]. When knowledge presentation is uncertain, it offers a theoretical framework to deal with that. Consequence of fuzzy logic develops from the fact that even human approaches of reasoning are suitable in nature. While classical logic deals with reasoning approaches, those provide exact analysis and formulations. Some significant features of fuzzy logic are:
ï¶ Exact reasoning is observed as a preventive case of approximate reasoning.
ï¶ The whole thing is a matter of degree.
ï¶ All logical method can be fuzzified.
Hence, it becomes clear that Fuzzy Logic is different from the classical logical system in detail and character e.g. accuracy, modifiers, probabilities, quantifiers and predicate.
...(download the rest of the essay above)