A NEW IMAGE EDGE DETECTION METHOD USING QUALITY-BASED CLUSTERING Bijay Neupane, Zeyar Aung, and Wei Lee Woon Computing and Information Science Masdar Institute of Science and Technology Masdar City, Abu Dhabi, UAE email: (bneupane, zaung, wwoon)@masdar.ac.ae ABSTRACT Due to the various limitations of existing edge detection methods, finding better algorithms for edge detection is still an active area of research. Many edge detection approaches have been proposed in the literature but in most cases, the basic approach is to search for abrupt change in color, in- tensity or other properties. Unfortunately, in many cases, images are corrupted with different types of noise which might cause sharp changes in some of these properties. In this paper, we propose a new method for edge detection which uses k-means clustering, and where different prop- erties of image pixels were used as features. We analyze the quality of the different clusterings obtained using different k values (i.e., the predefined number of clusters) in order to choose the best number of clusters. The advantage of this approach is that it shows higher noise resistance compared to existing approaches. The performance of our method is compared against those of other methods by using im- ages corrupted with various levels of “salt and pepper” and Gaussian noise. It is observed that the proposed method displayed superior noise resilience. KEY WORDS Edge detection, variance, entropy, gradient, busyness, sil- houette analysis. 1 Introduction Image edges contain useful information, which is very im- portant in image processing, machine vision, and pattern recognition mainly in the areas of feature extraction and detection. An edge is the region in the image where there is a sharp change in color intensity, discontinuities in depth and other properties. An edge may represent two differ- ent surfaces of the object or a boundary between light and shadow falling on a surface. Since none of the existing methods can produce the best results for all types of images for all types/levels of noises, finding better methods for edge detection is still an active area of research. Natural images are prone to noise and artifacts. Salt and pepper noise is a form of noise typically seen on im- ages. It is typically manifested as randomly occurring white and black pixels. Salt and pepper noise creeps into images in situations where quick transients, such as faulty switching, take place. On the other hand, White noise is additive in nature where the each pixel in the image is mod- ified via the addition of a value drawn from a Gaussian dis- tribution. To test the generality of the results, the proposed edge detection algorithm was tested on images containing both these types of noise. A large number of studies have been published in the field of image edge detection, which attests to its impor- tance within the field of image processing. Many edge de- tection algorithms have been proposed, each of which has its own strengths and weaknesses; for this reason, hitherto there does not appear to be a single “best” edge detector. A good edge detector should be able to detect the edge for any type of image and should show higher resistance to noise. Examples of approaches to edge detection include al- gorithms such as the Sobel, Prewitt and Roberts edge de- tectors which are based on the first order derivative of the pixel intensities. The Laplacian-of-Gaussian (LoG) edge detector is another popular technique, using instead the sec- ond order differential operators to detect the location of edges [1]. However, all of these algorithms tend to be sen- sitive to noise, which is an intrinsically high frequency phe- nomenon. To solve this problem the Canny edge detector was proposed, which combines a smoothing function with zero crossing based edge detection [2]. Although it is more resilient to noise than the previously mentioned algorithms, its performance is still not satisfactory when the noise level is high. There are many situations where sharp changes in color intensity do not correspond to object boundaries like surface marking, recording noise and uneven lighting con- ditions [3]. In this paper we propose a clustering based tech- nique for enhancing the performance of edge detection al- gorithms and providing better resistance to noise. This is achieved by filtering out outliers in the images and only identifying real object boundaries as edges. The proposed algorithm uses the variance, entropy, gradient, and busy- ness of each image pixel as a feature vector and employs the widely used k-means clustering algorithm on the pixel feature vectors in order to detect edge pixels. A further challenge is the determination of a suitable value of k (i.e., the number of clusters). Many of the previous papers on edge detection do not explain how this choice was made. Proceedings of the IASTED International Conference July 3 - 5, 2012 Banff, Canada Visualization, Imaging and Image Processing (VIIP 2012) DOI: 10.2316/P.2012.782-053 20