Image Mosaicing Using Binary Edge Detection - Semantic Scholar

Proceedings of the International conference on Computing Technology and Information Management, Dubai, UAE, 2014

Image Mosaicing Using Binary Edge Detection Inad Aljarrah1, Abdullah Al-Amareen2, Abdelrahman Idries2, Osama Al-Khaleel1 1

Computer Engineering Department, Jordan University of Sciences and Technology. Irbid, Jordan 1

2

{inad,oda}@just.edu.jo {amalamareen12,aaidraisalsouqi11}@cit.just.edu.jo

ABSTRACT Image mosaicing (panoramic imaging) is a key concept in image processing field. Essentially, it works by combining two or more images by stitching them together so that they look as one image. The main issue in image mosaicing is the alignment between the corresponding points so that overlapping regions between combined images are not repeated. The aim of images mosaicing research is concentrated on solving problems arising through different image mosaicing techniques. In this paper, a new technique that uses the edge profile of images to determine the corresponding points, then the overlapping region and the connecting curve which will stitch the two images to create the resulting mosaic image is determined. The approach is simple and easy to implement but still delivers promising results.

KEYWORDS Image Mosaicing, Panorama, Corresponding Points, Edge Detection, Image Matching.

1 INTRODUCTION Image mosaicing or panoramic imaging is an important concept in image processing field. The first and essential step of image mosaicing is the stereo pairs matching, which aims to find pairs of points that refer to the same vision in two images of the same scene taken from different views. After finding these points, the search about correlation between points is carried out. The correlation determines the overlapping region between the two images, and then stitching the two images in one image is performed. So image mosaicing goes through four steps, which are feature extraction, matching features to find corresponding points, establishing the homography

ISBN: 978-0-9891305-5-4 ©2014 SDIWC

between those points, and finally wrapping the images in a single image. Stereo pair matching is an important operation in any image processing approach that needs to find the similar points as vision between two images or more. In other words, image alignment aims at extracting stereo pairs between images. There are many methods to implement image alignment; the main ones are correlation that uses the image pixels values directly [1], image alignment that uses the frequency domain coefficients of images [2], edge and corners detection [3]; and features description that use high level features and descriptors to match the images [4]. Image homography is used to find the best transformation between the input images, this information is used to eliminate overlapping areas between those images. Image blending or wrapping is the last step in image mosaicing. After finding the matching points and overlapping region between images the stitching of images is applied to merge those in one image that contains all content of images without any repetition. Edges are defined as the objects boundaries in image. The pixels differential values on these points are the highest in the image matrix. The edge detection is done by looking for the pixels that achieve high differences with their neighborhood. There are many methods that are used in edge detection. The main methods employ a gradient approximation on the neighboring pixels within a window of a given size like the Sobel and Roberts operators. The Sobel operator sums the gradient values along two directions of image Gx and Gy. There are other operators to sharpening an image and finding its edges, e.g. Prewitt Operator, Kirsch Compass Masks, Robinson Compass Masks and Laplacian Operators. In this paper, we present an image mosaicing algorithm that uses the edge detection and the correlations between pixels to find the matching points and the overlapping region between two images. Then

186

Proceedings of the International conference on Computing Technology and Information Management, Dubai, UAE, 2014 those images are stitched together to create the mosaic image. The remaining of this paper is prepared as follow, section 2 talks about previous related work to image mosaicing. Section 3 presents the proposed approach and methods that are used to implement the algorithm. Section 4 summarizes the results and presents some examples. Section 5 concludes the paper.

2 RELATED WORKS This section summarizes the work related to image mosaicing. Jigsaw image mosaic (JIM) used tiles of arbitrary shapes to stitching a picture [5]. This algorithm solved the arbitrary-shaped container and tiles of images by filling the containers with tiles. Wavelet transformation to implement image mosaicing was proposed in [6]. The work focus on having a worthy advertising image, the model deals with nonfixed features. Another new algorithm used the speed up rebuts features (SURF) method to extract features with best-bin-first (BBF) to match features between two images [7]. This algorithm is robust to image rotation and image brightness variations. Stratified matching algorithm is used in image mosaicing [8]. Geometric hashing approach is used to generate another image mosaicing algorithm in [9] to reduce features matching time from exponential complexity to polynomial complexity. Taylor series is used with image features detection in image mosaicing [10]. It is used for estimation of perspective transform and parameters. Corner detection is used to find the corresponding pointes between two images then find the overlapping area between two images then stitching them [3]. Neural network is used as Kohonen’s SelfOrganizing Map (KSOFM) [11]. It used the right half of left image as trainer and left half of right image as tester for neurons of neural network to find the overlap area between two images. This algorithm requires the same size for two images. In [12], a firm and robust technique for image mosaicing is proposed. The methodology used two methods; the first uses a multi resolution spot-based optical flow to detect feature correspondences to obtain the homography. The second technique is a new method to obtain homography from only three points by dividing the scene into multiple triangles

The first step in this new technique is to convert the two input images to binary. In coloured images, a pixel is represented by 24 bits (3 bytes). Each byte can have a value from 0 up to 255. In gray scale, the pixel value ranges from 0 to 255 because only 8 bits are used to represent a pixel. The simplest representation of an image is in the binary scale where a pixel is represented by a single bit (0 for black, 1 for white). An example that shows how an image looks like before and after being converted from gray scale into binary is depicted in Figure 1 and 2.

Figure 1: Gray scale image.

3 METHODOLOGY This paper presents a new technique for image mosaicing based on edge detection. The inputs are two gray or coloured images of resolution M*N and L*W.

ISBN: 978-0-9891305-5-4 ©2014 SDIWC

Figure 2: Binary scale image.

187

Proceedings of the International conference on Computing Technology and Information Management, Dubai, UAE, 2014 Each pixel value in the gray image is either set to 0 or 1 in order to obtain the binary image. If the value of the pixel in the gray scale is less than a threshold then it is set to 0. Otherwise, it is set to 1. If the images are coloured, then they are converted into gray and then into binary. It is known that processing on binary images is very simple comparing with processing on gray or coloured images. However, if an image is converted into binary some details are lost because the pixel values are scaled down to 0 or 1. In our technique, binary format would have enough details. Therefore, we convert the image into binary. Once the binary format of the images is obtained, an edge detection algorithm is applied on the binary images in order to detect the edges. We have used Soble filter. The output of the Sobel filter is shown in Figure 3.

window in the second image is compared with the first image. The overlapping region is determined based on the highest similarity achieved. The adopted window size is (L/4)*(W/10) where L and W are the length and width of the second image.

3.2 The Similarity Correspondence Searching for corresponding points starts by aligning the (L/4)*(W/10) window to be centred at the beginning of the middle row of the second image, and we start to compare this window of the second image with the first image going left right, up down where each window in the first image is equal in the size to the fixed window in the second image. This operation is repeated and the result of each comparison (correlation matching) is recorded for the purpose of correspondence determination. For example, in the first turn let the window in the second image be called A and the window in the first image be called B. Now, each pixel in A will be multiplied by the corresponding one in B; if the two pixels from A and B are different the result will be zero while the result will be one if both pixels are equal to one. The sum of all the operations is tallied, and this operation is repeated and applied for all turns; after this step searching is done to find where the two images have highest similarity index, and the overlapping region between the two images is located accordingly. Input Images

Convert to

Gray Images

Sobel Filter

Matching through correlation

Figure 3: Image after applying Soble filter.

Edges will be used as a main parameter to find the overlapping regions and then the similarities between the two input images.

3.1 Finding The Overlapping Region Now the two images have edges only and each pixel has a value of 0 or 1. It is axiomatic that the overlapping region in the two images must have the same pixel values. However, in reality, there are differences in some pixel values after applying the edge detection on the images. Therefore, a fixed

ISBN: 978-0-9891305-5-4 ©2014 SDIWC

Edge Detection

Stitching images

Find correspond ing points

Figure 3: The proposed approach

4 RESULTS The proposed technique is implemented using MATLAB. MATLAB is a powerful environment to implement any image processing technique using the

188

Proceedings of the International conference on Computing Technology and Information Management, Dubai, UAE, 2014 built in functions that are available in its libraries. The input images are coloured ones. Thus, they are converted to gray level using the rgb2gray()function. The gray level output is then transformed into binary level using im2bw()function. After that the edges have been detected by using the edge()function. Many images have been tested. One example is shown by Figures 5 and 6 which depict the first and the second input images respectively.

Finally, figure 9 shows the output after stitching the two images.

Figure 9: Final result after mosaicing.

5 CONCLUSION

Figure 5: First Input Image

Figure 6: Second Input Image

This paper presents a new method for image mosaicing. The input images in our approach are converted to binary level and then the edges are detected using Soble filter. Overlapping regions are determined and the two images are stitched. The technique has been verified using Matlab. The proposed technique is not image size dependent and the two input images do not have to be of equal size.

Figure 7 and 8 show the two images after applying the edge detection.

7 REFERENCES

Figure 7: Edges of First Input.

Figure 8: Edges of Second Input

ISBN: 978-0-9891305-5-4 ©2014 SDIWC

[1]

J. Kim; Fessler, J.A. "Intensity-based image registration using robust correlation coefficients", Medical Imaging, IEEE Transactions on, page(s): 1430 - 1444 Volume: 23, Issue: 11, Nov. 2004.

[2]

B.S Reddy, B.N Chatterji, “An FFT-based technique for translation, rotation, and scale-invariant image registration”. IEEE Transactions on Image Processing 5(8), 1266–1271 (1996).

[3]

A. Rocha, R. Ferreira, and A. Campilho. “Image Mosaicing using Corner Detection”, in SIARP2000 - V Ibero-American Symposium on Pattern Recognition. 2000.

[4]

C. Lundie, “Feature Based Image Alignment”, October 25,2013.Avaialable:http://webpages.sdsmt.edu/~164062 1/CENG414_Cody_Lundie_Project2.pdf

[5]

J. Kim , F. Pellacini, “Jigsaw image mosaics”, Proceedings of the 29th Annual Conference on Computer Graphics and Interactive Techniques, Pages

189

Proceedings of the International conference on Computing Technology and Information Management, Dubai, UAE, 2014 657-664, SIGGRAPH 2002, San Antonio, Texas, USA, July 23-26, 2002. [6]

S.J. Banarase and M.R. Banwaskar, “Multiresolution Image Mosiacing Applications”, IJARCSSE, Volume 3, Issue 4, April 2013.

[7]

Y.Murali, V.mahesh, "Image Mosaic Using Speeded Up Robust Feature Detection", International Journal of Advanced Research in Electronics and Communication Engineering (IJARECE) Volume 1, Issue 3, September 2012, pp: 40-45.

[8]

Y. Kanazawa, K. Kanatani, "Image mosaicing by stratified matching", Image and Vision Computing, Volume 22, Issue 2, 1 February 2004, Pages 93-103.

[9]

U Bhosle, S Chaudhuri, Sd Roy, “A fast method for image mosaicing using geometric hashing”, IETE Journal of Research, 48(3-4), 317-324, 2002.

[10]

S.-H. Cho, Y.-K. Chung, and J. Y. Lee, "Automatic image mosaic system using image feature detection and taylor series," in Proceedings of VIIth Digital Image Computing: Techniques and Applications. Sydney, Dec. 2003, pp. 549-556.

[11]

S. Sharma, H. Tuli, S. Nagar, T. Dhir, S. Tayal, “Using Self-Organizing Neural Network for ImageMosaicing”, Advanced application of electrical engineering. Houston, 2009. pp. 76-80

[12]

N. Chiba, H. Kano, M. Minoh, M. Yasuda, M. Osumi, “Feature-Based Image Mosaicing”, IEICE, No.10, pp.1581-1589, 2000.

ISBN: 978-0-9891305-5-4 ©2014 SDIWC

190