Image Segmentation based on Connected ...

3 downloads 2729 Views 709KB Size Report
International Journal of Advances in Science and Technology,. Vol. .... information in a pixel context, it was possible to adapt it to varying image characteristics i.e. variation in color .... He has completed his Bachelors in Computer Science and.
International Journal of Advances in Science and Technology, Vol. 2, No.3, 2011

Image Segmentation based on Connected Component Analysis V.N.Santosh Kumar.Kuppili1, Prof.Dhanaraj Cheelu2, Prof.M.Rajasekhara Babu3,Dr. P.Venkata Krishna4 1

SCSE, VIT University, Vellore, Tamil Nadu, India Email id : [email protected]

2

SCSE, VIT University, Vellore, Tamil Nadu, India Email id : [email protected]

3

SCSE, VIT University, Vellore, Tamil Nadu, India Email id : [email protected]

4

SCSE, VIT University, Vellore, Tamil Nadu, India Email id : [email protected]

Abstract Character extraction from an image is one of the hot topics in the image processing area. There are many sub tasks that are involved in the character extraction. Some of the character extraction modules include binarization, segmentation and clustering. This paper discusses about various methodologies and algorithms for segmenting the text present in 2D images.

Keywords: Image Processing, Character Extraction, binarization, segmentation 1. Introduction Text extraction from an image is an application of pattern recognition for image processing that detects and identifies the text present in the image. Text Detection or recognition is a major milestone for processing digital images for detecting, classifying and tracking targeted images. Various methods exist for segmenting the characters present in an image. Connected Component analysis is one of the approaches used as an input to segmenting the characters present in an image. This approach needs all the adjacent pixels of a single character to be attached. Dilation helps the characters to stay more connected.

Figure 1. Original image and the corresponding dilation This paper talks about one of the most crucial steps in the character detection i.e. segmentation. The traditional approach would be to get the image in the digital form and extract the boundaries for each and every character that is present in the image.

Special Issue

Page 36 of 82

ISSN 2229 5216

International Journal of Advances in Science and Technology

2. Architecture The general format for character extraction and detection is as follows:

Figure 2. Flowchart of the proposed character detection methodology The architecture in Figure 2 specifies the order in which the characters in an image can be extracted and detected. The connected component module specifies the binarization and dilation to identify the contours of every character that is present in an image.

3. Segmentation Methods Separation of image into objects and background and isolating objects one from other by enhancing the representation of visual boundaries is known as Image Segmentation. It is always one of the primary and crucial step for object detection, recognition and identification. The result of image segmentation is the set of regions that are distributed collectively over the image or set of contours extracted from the image. Issues involved in the segmentation process include performance, input image and impact on image analysis system. Segmentations don’t worry about the type of characters extracted from the image. They, however, can be labeled in the post processing phase of the Figure 1. Several techniques and algorithms have been developed for image segmentation. There does not exist any generalized solution for the image segmentation problem as it depends on the domain knowledge and a wide variety of perspectives needs to be reviewed[1]. This paper reviews those methods that are based on finding object regions from gray-level images and possible for color images also. Some approaches that are reviewed in this paper are as follows – Edge Detection Methods, Histogram Based Methods, Tree/ Graph Based Methods, Region Splitting Methods, Region Growing Methods, Model Based Segmentation, Graph Partitioning Methods, Watershed Transformation, Multi-scale Segmentation, Probabilistic and Bayesian Approaches.

3.1 Edge Detection Methods Edge Detection methods are well explored in image processing. An edge is something where we find a sudden change in intensity from its adjacent pixels. The edges obtained from various edge

Special Issue

Page 37 of 82

ISSN 2229 5216

International Journal of Advances in Science and Technology

detection approaches are discontinuous. However, the edges can be made connected if the discontinuity is within a particular threshold. The most simple and accurate edge detection is Canny edge detection[2] and it can be integrated into large number of object recognition algorithms used in computer vision and image processing applications. Although, an image consists of noise and quantization errors, the success of a successful edge detector depends on the true location of edges. As the noise usually becomes the high frequency component, low pass filters are used as smoothing filters. The amount of smoothing depends on smoothing operator.

(a)

(b)

(c) (d) Figure 3. Edge detection Algorithms – Canny, Prewitt and Sobel Edge detection methods are used to identify the edges of each and every object in the image irrespective of their identity.

3.2 Histogram Based Methods In this technique, a histogram is computed for the whole image and the peaks and valleys are noted. The histogram is computed based on color or intensity as the measure. Recursive application of histogram seeking methods to clusters in the image is refinement to develop much smaller clusters. This is repeated until no more clusters are formed[3]. These methods are very efficient as they typically require only one pass through the pixels and limitation in identification of significant peaks and valleys in the image.

3.3 Graph Based Methods In this technique, Cho and Meer[4] proposed a new approach for segmentation, which is derived from the consensus of a set of different segmentation outputs on an input image. Instead of the statistics characterizing the spatial structure of the immediate neighborhood of a pixel, the collected statistics are used for determining local homoge neity for every pair of adjacent pixels. From the amalgamation of these initial segmentations, a co -occurrence probability is derived for every adjacent pixel pair which captures global information at the pixel or local level. The final segmentation is obtained by the co-occurrence probability.

3.4 Region Splitting Methods This is a divide and conquer method where the image is broken down into a set of disjoint regions which are similar within themselves. The approach is as follows:

Special Issue

Page 38 of 82

ISSN 2229 5216

International Journal of Advances in Science and Technology

    

Initially select the whole image as ROI. Next narrow down to a particular ROI within the image which satisfy some similarity constraint. If TRUE then ROI corresponds to a region in the image If FALSE then narrow down the selected ROI into more regions with each region as single ROI. Process continues until either no further splitting occurs or ROI is just one pixel in size.

(a) (b) Figure 4. (a) displays the original image and (b) displays the thinned image The result of the divide part includes many neighboring regions that have similar properties. A merging process is used which compares adjacent regions and merges them if necessary.

3.5 Region Growing Methods This is opposite to split and merge approach where a small region, called a seed pixel, is selected initially and regions are identified with pixels of similar properties. The procedure is as follows:  An initial set of small areas are selected and searched with adjacent pixels for si milar properties.  An arbitrary seed pixel is used to start up identifying the regions with similar properties  Addition of neighboring pixels leads to growing of the region.  When no adjacent pixel with similar property is detected, then that region is consi dered as a single cluster and another arbitrary seed pixel is chosen in the non -cluster area to find the other clusters.  The above process is repeated until all the pixels in the image fall under some cluster or belongs to some region.

(a) (b) (c) Figure 5. (a) displays the original image, (b) displays some access point and (c) displays the segmented image The feature mentioned above can be any feature such as intensity of pixels and intensity mean in an image. A variant of the above technique [5] is based on pixel intensities where mean and scatter of region and intensity of pixel are used as a test schematic. The distance threshold is used to either accept or reject a particular pixel from the cluster region.

Special Issue

Page 39 of 82

ISSN 2229 5216

International Journal of Advances in Science and Technology

3.6 Model Based Segmentation The basic assumption of this approach is that the structures of the objects/ regions will follow repetitive geometry. Hence, one can seek for a probabilistic model for t he explanation of variation in the shape of the object and impose the constraints while segmenting the image. This type of tasks involves  Registration of training samples to a standard pose  Probabilistic representation of variation of registered sample  Statistical inference between the model and the image. State of the art methods in the literature for knowledge-based segmentation including active shape and appearance models, active contours and deformable templates and level-set based methods.

3.7 Graph Partitioning Method In this technique, the image being segmented is modeled as weighted undirected graph. Each pixel is considered as a node in the graph and an edge exists between every pair of pixels. The similarity between the pixels forms the weight of an edge. The image is now partitioned into segments by removing the edges of the graph connecting them. The optimal partitioning of a graph is the one that minimizes the weights of the edges that were removed from the segments. This method called “Normalized Cuts” was proposed by Shi and Malik[6] which seeks the minimize the normalized cut, which is the ratio of the cut to all of the edges in the set.

3.8 Watershed Transformations The gradient magnitude of an image which acts as a topological surface is c onsidered to isolate the segments. The region boundaries are obtained from the pixels having the highest Gradient Magnitude Intensities. Water placed on any pixel enclosed by a common watershed line flows downhill to a common local intensity minima and thi s drain to a common minimum form a catchment basin which represents the regions. Brequelaire and Brun[7] have demonstrated about this transformation.

3.9 Multi-scale Segmentation Multi-scale segmentation or space-scale segmentation is a general framework for signal and image segmentation based on computation of image descriptors at multiple scales of smoothing. Witkin's [8][9] included the notion that with one scale parameter controlling the scale of segmentation, a one-dimensional signal could be unambiguously segmented into regions. Lindeberg [10] explored the problem of linking local extrema and saddle points over scales, and proposed an image representation called the scale -space primal sketch which makes explicit the relations between structures at different scales, and also makes explicit which image features are stable over large ranges of scale including locally appropriate scales for those. Bergholm [11] proposed to detect edges at coarse scales in scale -space and then trace them back to finer scales with manual choice of both the coarse detection scale and the fine localization scale.

3.10 Probabilistic and Bayesian Approaches In this technique, segmentation is initially done based on location of the intensities of each pixel and its neighbors in the co-occurrence matrix. Each pixel is then associated with a tuple which specifies whether it lies within a given region or a boundary pixel. This tentative segmentation was then refined. The algorithm is less effective if the clusters in the co-occurrence space have substantial overlap due to the imposition of local consistency. Since the techniques use global

Special Issue

Page 40 of 82

ISSN 2229 5216

International Journal of Advances in Science and Technology

information in a pixel context, it was possible to adapt it to varying image characteristics i.e. variation in color and texture.

3.11 Other Segmentation Approaches Campbell et al.[12] proposed a automatic segmentation and also classification method for any background images using neural networks. This technique allows images to be segmented using Self-Organizing Feature Maps (SOFM) based on texture and color information of the objects. They consist of 64x64 nodes for best segmentation. A few set of features are then extracted from each region. The features include average color, position, size, rotation, texture and shape. Now, classification is performed using a Multi Layer Perceptron with 28 input nodes and 11 output nodes. The training is performed on multiple regions and testing is done on a independent set of samples. Ng[13] describes an extension into the conventional k-means algorithm with the modification of the splitting rule in order to control the number of cluster members. Adding of suitable constraints into the mathematical program formulation, the author developed an approach that allows the use of k-means paradigm to efficiently cluster data sets with each cluster having fixed number of elements. Julinda Gllavata et al.[14] describes a histogram based segmentation for the objects in the image after subtracting them from the background. Edge detection and horizontal projectio n and text detection via local thresholding are the two major steps involved in extracting the objects from an image. Boiangiu et al.[15] describes the algorithm by way of determining for each character its bounded rectangle. It aims to determine the skew of the image around the main axis at angles of 0 o,90 o,180 o and 270 o. Huy et al.[16] describes strokes as a primary measure to segment the objects from the image. This method generally classifies non -background regions as objects in an image. The variance of the width of the objects is a measure of the amount of a variation of set of values. Variance = where, N is the number of continuous boundary pixels. G.Vamvakas et al.[17] and M.Sarfraz et al.[18] have demonstrated a hierarchy of segmentation from line, word and character. Initially, horizontal projection is applied to segment each and every line at various inclinations. Then the lines are made straight to give the vertical projections. Depending upon the thresholds used, the words and finally the characters can be extracted from the image. Kwangsoo Hahn, Youngjoon Han, Hernsoo Hahn [19] have proposed a new ellipse detection scheme using a modified Randomized Hough Transform (RHT) to use line segments. It detects line segments in an edge image and selects every pair to test whether they are pertained to the same ellipse of not using the RHT. If they pass the test, then they are merged. Since the proposed algorithm uses line segments, it reduces the computation time of the RHT significantly, and detects all ellipses included in an image without missing. The experimental results have shown that its performance is more prominent in detection of ellipses when they are overlapped and partially occluded.

4. Conclusion This paper discusses the importance of segmentation in image analysis and various methods of segmentation whose evaluation is an important field of study. Segmentation method depends heavily on the type of images[20] that are being used and these are adhoc in nature. An

Special Issue

Page 41 of 82

ISSN 2229 5216

International Journal of Advances in Science and Technology

evaluation of object segmentation algorithms for robot vision is given in [21]. The main objective of these algorithms is precisely segment the image without under - and oversegmentation. Kitchen and Rosenfeld[22] discussed various related to under- and oversegmentation of images. The key fact that helps segmentation algorithms go inaccurate is the lack of usage of context information in these algorithms.

5. References [1] T.Pavlidis, Algorithms for graphics and Image Processing, Springer, Berlin, 1982. [2] J.Canny, 1986, “A Computational approach to Edge Detection”, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. PAMI-8, pp.679-698 [3] Linda G. Shapiro and George C. Stockman(2001): “Computer Vision”, pp 279-325, New Jersy, Prentice Hall, ISBN 0-13-030796-3 [4] K.Cho and P.Meer, Image Segmentation from consensus information, Computer Vision and Image Understanding, vol. 68, no. 1, pp. 72-89, 1997 [5] R.M.Haralick and L.G.Shapiro, Survey-image Segmentation Techniques, Computer Vision Graphics ans Image Processing, vol. 29, pp. 100-132, 1985 [6] Jianbo Shi and Jitendra Malik(1997): “Normalized Cuts and Image Segmentation”, IEEE Conference on Computer Vision and Pattern Recognition, pp 731-737 [7] J.P.Braquelaire and L.Brun, “Image Segmentation with topological maps and inter-pixel representation, Journal of Visual Communication and Image Representation, vol. 9, no. 1, pp. 6279, 1998. [8] Witkin, A.P. “Scale-space filtering”, Proceedings of the 8th International Joint Conference on Artificial Intelligence, Karlsruhe, Germany, 1019-1022, 1983. [9] A.Witkin, “Scale-space Filtering: A new approach to multi-scale description,” Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, vol. 9, 1984, pp 150153. [10] Lindeberg.T, “Detecting salient blob-like image structures and their scales with a scalespaceprimal sketch: A method for focus-of-attention”, International Journal of Computer Vision,1993, pp.283-318 [11] Bergholm.f, “Edge Focusing”, IEEE Transactions on Pattern Analysis and Machine Intelligence,pages: 726-741, 1987. [12] N.W.Campbell, B.T.Thomas, T.Troscianko, “Automatic Segmentation and Classification of outdoor games using neural networks,” International Journal ofNeural Systems, pp.137-144, 1997 [13] M.K. Ng, “A Note on Constrained K-Means algorithm”, International conference on Pattern Recognition, pp. 515-519, 2000 [14] Julinda Gllavata, Ralph Ewerth, Bernd Freisleben, “Finding Text in Images via Local Thresholding”, Proceedings of the 3rd IEEE International Conference Symposium on Signal Processing and Information Technology, 2003. [15] Boiangiu Costin-Anton, Bogdan Raducanu, Andrei-Cristian Spataru, “High Precision Orientation and Skew Detection for Texts in Scanned Documents”, 5th IEEE Conference on Intelligent Computer Communication and Processing, 2009. [16] Huy Phat Le, Toan Nguyen Dinh, Soo Hyung, Kim, Guee Sang Lee, “Text Detection in Binarized Text Images of Korean signboard by Stroke Width Feature”, 10th IEEE International Conference on Computer and Information Technology, 2010. [17] Vamvakas, G.; Gatos, B.; Stamatopoulos, N.; Perantonis, S.J., “A Complete Optical Character Recognition Methodology for Historical Documents”, The Eighth International workshop on Document Analysis Systems,2008 [18] Sarfraz, M.; Mahmoud, S.A.; Rasheed, Z., “On Skew Estimation and Correction of Text”, Computer Graphics, Imaging and Visualization, 2007. [19] Kwangsoo Hahn, Youngjoon Han, Hernsoo Hahn, “Ellipse Detection Using a Randomized Hough Transform based on Edge Segment Merging Scheme”, Proceedings of the 6th WSEAS International Conference on Signal Processing, Robotics and automation, 2007, pp. 1-6 [20] G.N.Srinivasan, Dr. Shobha.G, “Segmentation Techniques for Target Recognition”, WSEAS Transactions on Computers, Vol 7, pp. 1556-1563, 2008.

Special Issue

Page 42 of 82

ISSN 2229 5216

International Journal of Advances in Science and Technology

[21] Stefan Florczyk, “Evaluation of Object Segmentation Algorithms as Components of Autonomous Video Based Robot Systems”, Proceedings of the WSEAS International Conferences, pp. 15911596, 2002. [22] L.Kitchen, A.Rosenfeld, “Scene Analysis using Region-based Constraint Filtering”, International Conference on Pattern Recognition, vol 17,no 2, pp. 189-203, 1984.

Authors Profile V.N.Santosh Kumar.Kuppili,was born in Visakhapatnam, AP in the month of June. He has completed his Bachelors in Computer Science and Engineering from JawaharlalNehru Technological University and pursuing his Masters degree at School of Computer Science and Engineering, VIT University, Vellore, Tamilnadu, India. His areas of interest include Image Processing and Pattern Recognition. He has been working in this field from the past six months and has acquired knowledge in various sub-areas of Image processing and Pattern recognition.

Dhanaraj Cheelu, He is working as Senior Assistant Professor at School of Computing Science and Engineering, VIT University, Vellore. He took his B.Tech degree in computer science and engineering from SRKREC Bhimavaram, Andhrapraesh and his M.Tech in computer science and engineering from National Institute of Technology, Calicut, kerala. He has four years software industry experience. Currently he is pursuing his research in Network sercurity. His areas of interest include Network Security, image processing and theoretical computer science.

M.Rajasekhara Babu, He is a Senior faculty member at School of Computing Sciences, VIT University, Vellore, India. He has completed his Bachelors in Electronics and Communication Engineering from Sri Venkateswara University, Tirupathi, India and took his Masters in Computer Science and Engineering from Regional Engineering College(NIT), Calicut and presently pursuing research at School of Computing Sciences, VIT University, Vellore in the area of “Analysis of power schema's for Compiler Optimization Techniques on heterogeneous multi-core architectures”. He has produced number of National and International papers and articles in reputed Journals and Conferences. His areas of interest include Multi core Architectures, Compilers. Dr. P.Venkata Krishna is with School of Computing Science and Engineering, VIT University, Vellore, India. He received his B. Tech in Electronics and Communication Engineering from Sri Venkateswara University, Tirupathi, India, M. Tech in Computer Science & Engineering from REC, Calicut, India and he received his Ph.D. from VIT University, Vellore, India. His research interests include Mobile and wireless systems and Grid Computing. He has published several papers in national and international refereed journals and conferences. He is the editor for 2003, 2004 and 2006 ObCom International conference proceedings and he is founding member for ObCom International Conference. He has authored few books on Computer Networks and Programming Languages. He is the reviewer for many reputed journals like IET Journal on Communications, International Journal of Security and Networks (IJSN) etc.

Special Issue

Page 43 of 82

ISSN 2229 5216