You are on page 1of 12

Edge detection in noisy images by neuro-fuzzy processing M.

Emin Yksel
a , a,

[Author vitae]

Digital Signal and Image Processing Laboratory, Department of Electrical and Electronics Engineering, Erciyes University, Kayseri, 38039, Turkey Received 21 December 2005. Available online 29 March 2006.

Abstract
A novel neuro-fuzzy (NF) operator for edge detection in digital images corrupted by impulse noise is presented. The proposed operator is constructed by combining a desired number of NF subdetectors with a postprocessor. Each NF subdetector in the structure evaluates a different pixel neighborhood relation. Hence, the number of NF subdetectors in the structure may be varied to obtain the desired edge detection performance. Internal parameters of the NF subdetectors are adaptively optimized by training by using simple artificial training images. The performance of the proposed edge detector is evaluated on different test images and compared with popular edge detectors from the literature. Simulation results indicate that the proposed NF operator outperforms competing edge detectors and offers superior performance in edge detection in digital images corrupted by impulse noise. Keywords: Neuro-fuzzy systems; Image processing; Edge detection

Article Outline
1. Introduction 2. Method 2.1. The proposed neuro-fuzzy operator 2.2. The neuro-fuzzy subdetectors 2.3. The postprocessor 2.4. Training of the neuro-fuzzy subdetectors 2.5. Processing of the noisy input image 3. Results 4. Discussion and conclusion References Vitae

1. Introduction

Edges in a digital image provide important information about the objects contained within the image since they constitute the boundaries between the objects in the image. Edge detection is a frequently performed operation in many image processing applications because it is usually the first operation that is performed before other image processing tasks such as image segmentation, boundary detection, object recognition and classification, image registration, and so on. Consequently, the success of these subsequent image processing tasks are strictly dependent on the performance of the edge detection operation. The image intensity shows sudden changes at edges. Therefore, edge detection usually involves the calculation of the derivative of the image intensity function at a given pixel location. If the magnitude of the derivative of the image intensity function is relatively high at a given pixel location, then the pixel at that image location is classified as an edge pixel. The most important factor decreasing the performance of edge detection is the noise. Unfortunately, digital images are inevitably degraded by noise during image acquisition and/or transmission due to a number of imperfections encountered in imaging processes and/or communication channels. Most edge detection operators are based on the assumption that images contain large homogeneous regions separated by clear boundaries. However, this assumption loses its validity if the image is corrupted by noise. Therefore, majority of the edge detection operators require a prefiltering of the noisy image by using an appropriate noise filter before edge detection is performed. In this case, however, the performance of the edge detection operation becomes strictly dependent on the performance of the noise filter. Moreover, the complexity of the system and the processing time are considerably increased. A number of methods for edge detection implementing different approaches to the digital calculation of the derivative of the image intensity function are available in the literature. The classical methods [1] and [2] such as the Sobel, Prewitt and Kirsch detectors calculate the first directional derivative to determine the locations of the edges. These detectors are simple to implement but they are usually inaccurate and highly sensitive to noise. The zerocrossing edge detectors [1] and [2] use the second derivative along with the Laplacian operator. These detectors have fixed detection characteristics in all directions but they are very sensitive to noise too. The Gaussian edge detectors [2] reduce the undesirable negative effects of noise by smoothing the image before performing edge detection. Hence, they exhibit much superior performance over other operators especially in noisy conditions. The Canny detector, which is a Gaussian edge detector, is one of the most popular edge detectors in the literature and it has been widely used in many applications [3], [4] and [5]. Although the Gaussian detectors exhibit relatively better performance, they are computationally much more complex than classical derivative based edge detectors. Furthermore, their performances quickly decrease as the density of the corrupting noise increases. Therefore, a novel edge detector that is capable of extracting edges from digital images corrupted by noise is highly desirable. In the last few years, there has been a growing research interest in the applications of soft computing based techniques, such as neural networks and fuzzy logic systems, to various

problems in image processing [6], [7], [8], [9], [10], [11], [12], [13], [14] and [15]. This is due mainly to the fact that neuro-fuzzy (NF) systems are very suitable tools to deal with uncertainty encountered in the process of extracting useful information from noisy images since NF systems combine the ability of neural networks to learn from examples and the capability of fuzzy logic systems to model the uncertainty and imprecision. Hence, NF systems may be employed as powerful tools for edge detection provided that appropriate network topologies and training strategies are chosen. In our recent work, we have shown that a NF system may be utilized to remove impulse noise from digital images [16] and [17]; to construct highly efficient hybrid filters for restoring noisy digital images while preserving edges, lines and other useful information within the image [18]; to detect noise for guiding switching noise filters and reducing their undesirable blurring effects [19] and [20]; and to improve noise suppression and detail preservation performances of image filters [21]. In addition to these successful applications, we have also shown in a preliminary report [22] that NF systems may be employed for efficient detection of edges in noisy digital images. In this paper, we extend our preliminary research on edge detection and present a novel NF method for edge detection in digital images corrupted by impulse noise. In the proposed method, the edges in the noisy image are directly determined by a NF network without needing a prefiltering of the noisy input image. The NF network consists of a desired number of subdetectors and a postprocessor. Each subdetector evaluates a different pixel neighborhood in the filtering window. The proposed NF edge detector is tested on popular images having different image properties and also compared with popular edge detectors from the literature. Experimental results show that the proposed NF edge detector exhibits much better performance than the competing operators and may efficiently be used for the detection of edges in digital images corrupted by impulse noise.

2. Method
2.1. The proposed neuro-fuzzy operator Fig. 1a shows the general structure of the proposed NF edge detection operator. The operator is constructed by combining a desired number of NF subdetectors with a postprocessor. All NF subdetectors in the structure operate on the same 3-by-3 pixel filtering window, which is shown in Fig. 1b. Each NF subdetector evaluates a different neighborhood relation between the center pixel of the filtering window and two of its neighbors. Some of the many possible neighborhood topologies are shown in Fig. 1c. The higher the number of NF subdetectors, the better the edge detection performance, but the higher the computational cost.

Full-size image (42K) Fig. 1. (a) The general structure of the proposed neuro-fuzzy edge detection operator. The pixels applied to the inputs of each NF subdetector in the structure are chosen so as to utilize the information from a different pixel neighborhood; (b) The filtering window of the operator; (c) Some of the possible pixel neighborhood topologies.

2.2. The neuro-fuzzy subdetectors Each NF subdetector is a first-order Sugeno type fuzzy inference system with 3-inputs and 1-output. The internal structures of the NF subdetectors are identical to each other. Each input has 3 generalized bell type membership functions and the output has a linear membership function. The inputoutput relationship of any of the NF subdetectors is as follows: Let X1, X2, X3 denote the inputs of the NF subdetector and Y denote its output. Each possible combination of inputs and their associated membership functions is represented by a rule in the rule base of the NF subdetector. Since the NF subdetector has 3 inputs and each input has 3 membership functions, the rule base contains a total of 27 (33) rules, which are as follows: 1. if (X1 is M11) and (X2 is M21) and (X3 is M31), then R1=F1(X1,X2,X3) 2. if (X1 is M11) and (X2 is M21) and (X3 is M32), then R2=F2(X1,X2,X3) 3. if (X1 is M11) and (X2 is M21) and (X3 is M33), then R3=F3(X1,X2,X3) 4. if (X1 is M11) and (X2 is M22) and (X3 is M31), then R4=F4(X1,X2,X3) 5. if (X1 is M11) and (X2 is M22) and (X3 is M32), then R5=F5(X1,X2,X3)

27. if (X1 is M13) and (X2 is M23) and (X3 is M33), then R27=F27(X1,X2,X3), where Mij denotes the jth membership function of the ith input, Rk denotes the output of the kth rule, and Fk denotes the kth output membership function. The input membership functions are generalized bell type

(1)

and the output membership functions are linear

(2)Fk(u1,u2,u3)=dk0+dk1u1+dk2u2+dk3u3with i=1,2,3; j=1,2,3 and k=1,,27. Here the parameters a, b, c and d are constants that characterize the shape of the membership functions. The optimal values of these parameters are determined by training, which will be discussed in detail later on. The output of the NF subdetector is the weighted average of the individual rule outputs. The weighting factor, wk, of each rule is calculated by evaluating the membership expressions in the antecedent of the rule. This is accomplished by first converting the input values to fuzzy membership values by utilizing the input membership functions and then applying the and operator to these membership values. The and operator corresponds to the multiplication of input membership values. Hence, the weighting factors of the rules are calculated as follows:

(3) Once the weighting factors are obtained, the output of the NF subdetector can be found by calculating the weighted average of the individual rule outputs

(4) Readers interested in the details of fuzzy systems may refer to an excellent book on this subject [23]. 2.3. The postprocessor The outputs of the NF subdetectors are fed to a postprocessor, which generates the final NF network output. The postprocessor actually calculates the average value of the subdetector outputs and compares this value with a threshold. The threshold value is the half of the available

dynamic range for the pixel luminance values. For 8-bit images, where the pixel values range between 0 and 255, the threshold value is 128. The inputoutput relationship of the postprocessor may be explained as follows. Let Y1,Y2,,YK represent the outputs of the NF subdetectors in the structure of the NF network, respectively, where K is the number of NF subdetectors used. The output of the postprocessor is calculated in two steps. In the first step, the average value of the individual NF subdetectors outputs are calculated

(5) In the second step, this value is converted to 0 (black) or 255 (white) by comparing it with the threshold

(6) where y(r,c) is the output of the postprocessor, which is also the output of the proposed NF edge detection operator. 2.4. Training of the neuro-fuzzy subdetectors The internal parameters of the proposed NF edge detector are optimized by training. Each NF subdetector in the structure is trained individually. Fig. 2 represents the setup used for training. Here, the parameters of the NF subdetectors under training are iteratively adjusted so that its output converges to the output of the ideal edge detector which, by definition, can correctly detect the locations of the edge pixels of the image fed to its input. The ideal edge detector is conceptual only and does not necessarily exist in reality. It is only the output of the ideal edge detector that is necessary for training, and this is represented by the target training image.

Full-size image (14K) Fig. 2. Training of the subdetectors.

Fig. 3 shows the images used for training. The images are 128-by-128 pixel artificial images that can easily be generated by computer. The image shown in Fig. 3a is the base training image. Each square box in this image has a size of 4-by-4 pixels and the 16 pixels contained within each box have the same luminance value. The image in Fig. 3b is the input training image and obtained by corrupting the base training image by impulse noise. The image in Fig. 3c is the target training image. It is a black and white image and its

black pixels indicate the locations of the true edges of the input training image. Hence it represents the output of the ideal edge detector for the input training images.

Full-size image (65K) Fig. 3. Training images: (a) base training image; (b) input training image; (c) target training image.

The images in Fig. 3b and c are employed as the input and the target (desired) images during training, respectively. The parameters of the subdetector under training are then tuned by using the LevenbergMarquardt optimization algorithm [23], [24] and [25] so as to minimize the learning error. 2.5. Processing of the noisy input image The overall procedure of applying the proposed NF edge detection operator to a noisy image may be summarized as follows: 1. A 3-by-3 pixel filtering window is slided on the image. The window is started from the upper-left corner of the image and moved sideways and progressively downwards in a raster scanning fashion. 2. For each window, the appropriate pixels of the filtering window representing different neighborhood topologies are fed to the corresponding NF subdetectors in the structure. Each NF subdetector individually generates an edge magnitude value for the center pixel of the filtering window by processing the luminance values of its input pixels. 3. The outputs of the NF subdetectors are fed to the postprocessor, which calculates the average value of the NF subdetector outputs and suitably truncates this value to 0 or 255. The value obtained at the output of the postprocessor is also the output value of the NF edge detector and represents the decision whether the center pixel of the filtering window is an edge pixel or not. 4. This procedure is repeated for all pixels of the noisy input image.

3. Results
The NF edge detector presented in the previous section is implemented and applied to a number of popular test images from the literature including the Bone, Lena, Peppers and Sail images, which are shown in Fig. 4. All test images are 8-bit grey level images.

Full-size image (240K) Fig. 4. Test images. From top to bottom: Bone, Lena, Peppers, Sail. In each pair, the image on the left is the original (noise-free) image and the image on the right is the noisy image used in the edge detection experiments.

The noisy images used in the edge detection experiments are obtained by corrupting the original test images by impulse noise with 20% noise density. The noisy test images are shown in Fig. 4 too. Two different implementations of the proposed NF edge detector are tested in experiments. The first implementation contained 2 NF subdetectors while the second implementation contained 8 NF subdetectors. The neighborhoods applied to the inputs of the NF subdetectors are chosen as different as possible especially when constructing the implementation with 8 subdetectors. In order to make comparisons, the same noisy images are also processed by using the Sobel and the Canny edge detectors. Fig. 5 shows the output images of all detectors for the noisy experimental images. It is observed from this figure that the performance of the Sobel detector is very poor. For all test images, its output images are severely degraded by noise. Most noise pulses are incorrectly detected as edges. Furthermore, only the edges with sharp intensity variation, as in the Lena image, are partially reflected to the output image while the edges with less intensity variation are completely lost in the noise.

Full-size image (241K) Fig. 5. Comparison of the proposed edge detector with the Sobel and the Canny edge detectors.

The Canny detector demonstrates a considerably better performance than the Sobel detector. It correctly detects most of the noise pulses, especially when they are located within relatively homogenous regions of the images, and these are not reflected to the output image as edges. This is observed in the Peppers and the Sail images. However, the effect of noise is still clearly visible as real edges are significantly distorted by the noise and many noise pixels are incorrectly detected as edges. Due to the Gaussian smoothing of the detector, the edges in the more detailed regions of the input image are not properly detected, as for the Lena image. This undesirable side effect of Gaussian smoothing is easily observed in the response of the Canny operator to the noisy Lena image, where most of the facial details are lost. In addition, the edges where the luminance variation is not very sharp are almost lost in the noise. This is especially observed in the Bone and the Sail images. On the other hand, both implementations of the proposed NF edge detector exhibit very good detection performance and successfully detect most of the edges in all images. It is observed that the implementation with 8 subdetectors performs significantly better than the two subdetector implementation. The effect of noise on the performance of the detector is much less compared to the Sobel and the Canny detectors. Object boundaries and other details in the images are reflected to the output image of the proposed detector much better than the other detectors. The edges in the more detailed regions of the input images are successfully extracted, as observed in the Lena image. In addition, the proposed detector shows a significantly better performance at detecting the edges where the luminance variation is not very sharp. The performance difference is very clear especially for the Bone and the Sail images.

4. Discussion and conclusion


A novel NF edge detector for digital images corrupted by impulse noise is presented. The advantages of the proposed NF edge detector may be summarized as follows: 1. It does not require the prefiltering of the noisy input image with a noise filter prior to edge detection. 2. It has a flexible structure. The number of NF subdetectors in the structure can be varied to obtain the desired performance. The higher the number of NF subdetectors included in the structure of the operator, the better the edge detection performance, but the higher the computational complexity. 3. It has a very simple structure. Its fundamental building block is a 3-input 1-output NF system. This greatly simplifies implementation. 4. It is trained by using very simple artificial images, which can easily be generated in a computer. It is concluded that the proposed edge detector can be used for efficient extraction of edges in digital images corrupted by impulse noise.

References
[1] S.E. Umbaugh, Computer vision and image processing, Prentice-Hall International Inc., Englewood Cliffs, NJ (1998). [2] R. Albovik, Handbook of image and video processing, Academic Press, New York (1998). [3] Linares I, Maersereau RM, Smith MJT. JPEG estimated spectrum adaptive post filter using image-adaptive Q-tables and Canny edge detectors. ISCAS-96. IEEE International Symposium on Circuits and Systems, Atlanta, GA, USA, 1996. pp. 72225. [4] Ali M, Clausi D. Using the Canny edge detector for feature extraction and enhancement of remote sensing images. IGARSS-01. IEEE International Symposium on Geoscience and Remote Sensing, Sydney, Australia, 2001. pp. 2298300. [5] Hongjian S, Ward R. Canny edge based image expansion. ISCAS-2002. IEEE International Symposium on Circuits and Systems, Scottsdale, AZ, USA, 2002. pp. 785 88. [6] Y.S. Choi and R. Krishnapuram, A robust approach to image enhancement based on fuzzy logic, IEEE Trans Image Processing 6 (1997) (6), pp. 808825. View Record in Scopus | Cited By in Scopus (88) [7] D.L. Pham and J.L. Prince, Adaptive fuzzy segmentation of magnetic resonance images, IEEE Trans Med Imaging 18 (1999) (9), pp. 737752. Full Text via CrossRef | View Record in Scopus | Cited By in Scopus (304) [8] C.-T. Lin, Y.-C. Lee and H.-C. Pu, Satellite sensor image classification using cascaded architecture of neural fuzzy network, IEEE Trans Geosci Remote Sensing 38 (2000) (2), pp. 10331043. View Record in Scopus | Cited By in Scopus (25) [9] D.C. Park and Y.J. Woo, Weighted centroid neural network for edge preserving image compression, IEEE Trans Neural Networks 12 (2001) (5), pp. 11341146. View Record in Scopus | Cited By in Scopus (36) [10] R. Carotenuto, G. Sabbi and M. Pappalardo, Spatial resolution enhancement of ultrasound images using neural networks, IEEE Trans Ultrasonics Ferroelectrics Frequency Control 49 (2002) (8), pp. 10391049. Full Text via CrossRef | View Record in Scopus | Cited By in Scopus (8) [11] S.-J. Lee, C.-S. Ouyang and S.-H. Du, A-neuro fuzzy approach for segmentation of human objects in image sequences, IEEE Trans Syst Man Cybernet (Part B) 33 (2003) (3), pp. 420437. View Record in Scopus | Cited By in Scopus (13)

[12] D.Q. Zhang and S.C. Chen, A novel kernelized fuzzy C-means algorithm with application in medical image segmentation, Artific Intell Med 32 (2004) (1), pp. 3750. Article | PDF (722 K) | View Record in Scopus | Cited By in Scopus (75) [13] M.K. Bashar, N. Ohnishi, T. Matsumoto, Y. Takeuchi, H. Kudo and K. Agusa, Image retrieval by pattern categorization using wavelet domain perceptual features with LVQ neural network, Pattern Recognition Lett 26 (2005) (15), pp. 23152335. Article | PDF (1457 K) | View Record in Scopus | Cited By in Scopus (5) [14] S.S. Park, K.K. Seo and D.S. Jang, Expert system based on artificial neural networks for content-based image retrieval, Expert Syst Appl 29 (2005) (3), pp. 589597. Article | PDF (608 K) | View Record in Scopus | Cited By in Scopus (8) [15] Y. Zhuge, J.K. Udupa and P.K. Saha, Vectorial scale-based fuzzy-connected image segmentation, Comput Vision Image Understanding 101 (2006) (3), pp. 177193. Article | PDF (618 K) | View Record in Scopus | Cited By in Scopus (18) [16] M.E. Yksel and A. Batrk, Efficient removal of impulse noise from highly corrupted digital images by a simple neuro-fuzzy operator, AE Int J Electron Commun 57 (2003) (3), pp. 214219. Full Text via CrossRef | View Record in Scopus | Cited By in Scopus (30) [17] M.E. Yksel and A. Batrk, A simple generalized neuro-fuzzy operator for efficient removal of impulse noise from highly corrupted digital images, AE Int J Electron Commun 59 (2005) (1), pp. 17. View Record in Scopus | Cited By in Scopus (10) [18] M.E. Yksel, A hybrid neuro-fuzzy filter for edge preserving restoration of images corrupted by impulse noise, IEEE Trans Image Processing 15 (2006) (4), pp. 928936. Full Text via CrossRef | View Record in Scopus | Cited By in Scopus (22) [19] M.E. Yksel and E. Bedok, A simple neuro-fuzzy impulse detector for efficient blur reduction of impulse noise removal operators for digital images, IEEE Trans Fuzzy Syst 12 (2004) (6), pp. 854865. Full Text via CrossRef | View Record in Scopus | Cited By in Scopus (30) [20] M.E. Yksel, A. Batrk and E. Bedok, Detail preserving restoration of impulse noise corrupted images by a switching median filter guided by a simple neuro-fuzzy network, EURASIP J Appl Signal Process 2004 (2004) (16), pp. 24512461. Full Text via CrossRef | View Record in Scopus | Cited By in Scopus (17) [21] M.E. Yksel, A simple neuro-fuzzy method for improving the performances of impulse noise filters for digital images, AE Int J Electron Commun 59 (2005) (8), pp. 463472. Article | PDF (483 K) | View Record in Scopus | Cited By in Scopus (3)

[22] M.E. Yksel and M.T. Yldrm, A simple neuro-fuzzy edge detector for digital images corrupted by impulse noise, AE Int J Electron Commun 58 (2004) (1), pp. 7175. [23] J.S.R. Jang, C.T. Sun and E. Mizutani, Neuro-fuzzy and soft computing, Prentice-Hall International Inc., Englewood Cliffs, NJ (1997). [24] K. Levenberg, A method for the solution of certain problems in least squares, Quan Appl Math 2 (1944), pp. 164168. MathSciNet [25]D.W. Marquardt, An algorithm for least squares estimation of nonlinear parameters, J Soc Ind Appl Math 31 (1963), pp. 431441. MathSciNet | Full Text via CrossRef Tel.: +90 352 4374901/32204; fax: +90 352 4375784.

Vitae
M. Emin Yksel received his B.Sc. degree in Electronics and Communications Engineering from the Technical University of Istanbul, Istanbul, Turkey, in July 1990. In February 1991, he joined as a research assistant the Department of Electronics Engineering, Erciyes University, Kayseri. He received his M.Sc. and Ph.D. degrees in Electronics Engineering from Erciyes University in February 1993 and September 1996, respectively. Between March and December 1995, he had been an academic visitor to Signal Processing Section, Department of Electrical Engineering, Imperial College, London, UK, where he conducted research on time varying parametric modeling of nonstationary signals. Currently, he is with the Department of Electronics Engineering of Erciyes University. His general research interests include signal processing, image processing, neural networks, fuzzy systems and applications of these techniques. AEU - International Journal of Electronics and Communications Volume 61, Issue 2, 1 February 2007, Pages 82-89

You might also like