A Parallel Approach to Combine SVM, Edge and Corner Detection Methods for Target Detection

Saeed Mirghasemi Department of Computer and IT Engineering Islamic Azad University Parand Branch Parand, Iran [email protected]

ABSTRACT: This paper presents a simple and effective approach for target detection in sea images. In this regard, three renowned detection methods which are SVM, edge and corner detectors have implemented. They have been applied on an image independently, and then a combination between detected points have been made by sliding a 20×20 matrix. To decrease the false positives and to increase the efficiency of final detection results, connected component labelling has been applied to the output of each detector to eliminate small connected components. Testing the proposed method on two different databases of sea target images shows its efficiency in target detection. Keywords: Sea Target Detection, Support Vector Machine, Corner Detection, Edge Detection, Color Features, Connected Component Labelling Received: 1 September 2012, Revised 28 September 2013, Accepted 4 October 2013 © 2013 DLINE. All rights reserved 1. Introduction Sea target detection methods through image processing, based on the type of images that have been used, are divided into four categories: a) RADAR images b) IR images c) Satellite images d) Visible images Of which the first and most considered one concerns the sea target detection based on radar imagery. The radar system is essential for detecting in adverse weather conditions such as fog, raining, snowing, etc [1]. Two significant kinds of radar images are SAR and ISAR images. Inverse synthetic aperture radar (ISAR) is an imaging technique which is used to produce medium/high resolution images (metric or sub-metric) of naval, aerial or ground targets. Most of researches in this area deal with radar imaging techniques rather than image processing techniques. One of the most considered parts of radar imagery are the SAR images. At present, there are two modes to detect ships in SAR images. One is a direct mode, detecting ships directly. The other is an indirect mode. That is, it first detects ship wakes, and then seeks ships around the wakes. In the direct mode, there are a lot of methods to detect ships, such as adaptive threshold way [2] [3], probability neural network (PNN) model method [4], double parameter constant false-alarm ratio (CFAR) detection method [4], [5], and fractal detection algorithm [6]. Ship wakes Journal of Multimedia Processing and Technologies Volume 4 Number 2 June 2013

47

detection methods mainly include Radon transform, Hough transform, mathematical morphology and wavelet analysis [7]. However, the wake of a slowly moving or nearly stationary ship is unlikely identifiable in a SAR image [8]. All the SAR-based methods expend largely but they have two serious drawbacks: a) They can only obtain target points, which cannot be used to recognize the targets. b) There are targets that are invisible to radar like wooden boats. The second group consists of detection methods based on IR images. The RADAR images are often undesirable for military applications because they reveal the location of the imaging system. So, researchers explored visible and infrared images of ships which are generally more consistent than RADAR images and for which it is easier to compensate for environmental effects [9]. The infrared system is employed to enhance vision in weak light conditions. To extract semantic objects from a scene, an a priori knowledge of the image type and object characteristics is necessary. The a priori knowledge for the infrared image segmentation is based on the fact that the object that has to be extracted has either a larger or smaller temperature than the environment, being characterized by transient elements such as edges and peaks. The work published in [10] is one of recent years papers that worked on IR images. The authors used PCA, Bayes classification and wavelet-denoising to classify the sea targets, but in several papers, the limitations and disadvantages of methods based on statistical analysis are pointed out [11]. IR images have three main problems [9]: e) Poor SNR-ratio f) Varied gray levels g) Naval ships or small boats have a homogenous stern side. The third group consists of methods that use visible satellite images for sea target detection [12] [13]. One part of the last group is satellite high resolution images. High-resolution images allow a more accurate and reliable discrimination of ships and the possibility of their classification. However, the detection technique for this case must deal with a larger level of nonhomogeneity and must take into account the intrinsic non-Gaussian nature of the back-scattered intensity [13]. Hu et.al [13] present a method based on cumulative projection curve (CPC) to estimate the number of ships of small size, which is only efficient on special images of stationary ships along coastline. One of a few research works that uses color feature, from the Lab color coordinate system, in this area is [12]. They have presented a definition on the degree of overlap between two clusters and developed an algorithm for calculating the overlap rate. Using this theory, they also have developed a new hierarchical cluster merging algorithm for image segmentation and apply it to the ship detection in high resolution image. The last group consists of methods that work on visible images. Most of these methods use grey-level features for target detection [11] [15]. One of the research works that is more superior to previous works in this group is [11]. Their work is based on calculating different chaos by obtaining largest Lyapunov exponent of target and sea background which is not appropriate for images that contain some low chaos object other than the target. Also the same authors have proposed the work in [15] based on the natural measure feature. Although the method’s results are considerable for some images but it still suffers from previously mentioned imperfection and needs analyzing several frames for exact results. In this group a few methods have been proposed for sea target detection in color images [16] [17] [18]. In [16] a new target-based color space based on a new combination of PSO and FCM is introduced. The authors in [17] and [18] use color features of the HSV color space to train SVM and a new kernel-based neural network for sea target detection respectively. 2. The Proposed Method The flowchart of the proposed is shown in Figure 1. The method attempts to propose a new simple combination between the detection results of three famous methods; Support Vector Machine (SVM), edge detection, and corner detection. While SVM tries to classify the body of a target, edge and corner detectors mostly aims the points located in the contour of it. Among these detection methods, the SVM is a supervised method and needs some pre-knowledge before classification. Sea targets in color images have been used as benchmark to evaluate the final detection results of the proposed method. Color features of the target and non-target are used to train the SVM. Therefore, the first step is dedicated to extraction of training sample. These samples are extracted from target and non-target parts of an image in the RGB color space as a 20 × 20 matrix. The actual nature of the samples is a 20 × 20 × 3 matrix, in which 3 indicates R, G and B color components of the image in the RGB color space. In the next step, we apply the edge detector which is the Canny edge detector, and the corner detector which has been proposed

48

Journal of Multimedia Processing and Technologies Volume 4 Number 2 June 2013

in [19], and the SVM classifier in a parallel manner on our test image. The outputs of every three steps are binary images with white parts consumed as target. The results of each detector has many false positives as depicted in Figure 2. A very effective way to reduce the false positives is eliminating small connected components. The Original Image in the RGB Color Space

Training Sample Extraction

Corner Detection

Edge Detection

Small Connected Components Elimination

Small Connected Components Elimination

SVM Detection

Small Connected Components Elimination

Perserving Common Points via Sliding a 20 × 20 Matrix Figure 1. The flowchart of the proposed method

Figure 2. Showing false positives after SVM, edge and corner detection. a) The original image. b) Corner detection. c) Edge detection. d) SVM detection Journal of Multimedia Processing and Technologies Volume 4 Number 2 June 2013

49

We have utilized the proposed method introduced in [20] to find separate connected areas, and remove the small ones. The question is, is it possible to find proper thresholds for the small connected components of each output which could effectively work in any images of different size or different shape of the target? We have three different threshold parameters for this goal in each detector which will be considered with more details in the next section. Now, we slide a 20 × 20 on the image obtained from SVM detector, and find the common detected points of all three methods. If in a window we have points detected as the target via all three methods, the window would be preserved without any changes. Otherwise, all the detected points would be treated as non-target. In other words, if in a window one of the detectors doesn’t have a detected point, all the detected points would be considered as non-target. There is a possibility that the sliding window cannot find any common points in the three outputs of the connectedcomponentelimination level. That is to say, after eliminating small connected components, one of the primary detection results would not have any useful points as the target. In this case the detection results would be invalid. Sea Figure 3 in which the corner detection results, and the output after omitting small connected components is depicted. As the image shows, due to bad lighting condition of the image, there are a few corners related to the target (Figure 3-b). Therefore, all the detected points related to the target have been eliminated after performing small connected components elimination level (Figure 3-c).

Figure 3. A case in which the elimination level after the detection level leaves no points related to the target. a) The original image. b) Corner detection results. c) Corner detection after elimination small connected components To remove this problem, first, the method recognizes which one of the detection methods has no valid points as target after elimination, and then uses its detection result before elimination to be used in sliding window procedure. 3. Experimental Results To evaluate the effectiveness of the method properly, it is tested on two different databases. The first database has been created by the authors for sea target detection purposes. i.e. the image are standards in framing rules, size of the target, quality of the image and background considerations, and the second database is composed of different images with different quality, size and background information from main stream movies and websites. It is logical that the results are more acceptable in the first database, but even the method works affectively in the second database. After applying the method on different images of the two databases, we set the thresholding parameters to eliminate small connected components in corner, edge and SVM detection results, 120, 50 and 220 pixels, respectively. These thresholds are applicable in both databases. Figure 4 shows different images of the first database along the outputs of each level toward the final detection. As observable in Figure 4, there are images in which none of the detection methods could hand in the exact results individually. Images no. 2 and 3 are among the images in which only the introduced combination has leaded to the proper final detection. Another preference of the method is its ability in finding the target even in images with bad lighting condition, which is very probable in real sea target images. As a case in points, the image no. 2 has sun as its background which has caused the similarity between the color components of the target and its background. Although, the final detection is possible with applying only the corner detection and the post elimination of small connected component, but the aim of this paper is to introduce a general method for sea target detection which is applicable in all sea images. The detection results of the second database has shown in Figure 5. Here again, we have images in which only the proposed combination is effective (image 1). Also, we have images with complicated backgrounds with acceptable detection results images 2, 3 and 4).

50

Journal of Multimedia Processing and Technologies Volume 4 Number 2 June 2013

(a)

(b)

(c)

(d)

(e)

(f )

(g)

(h)

(i) (1) (2) (3) (4) Figure 4. Detection results on images from the first database. a) The original image. b) Corner detection. c) Corner detection after elimination of small connected components. d) Edge detection. e) Edge detection after elimination of small connected components. f) SVM detection. g) SVM detection after elimination of small connected components. h) Detection results after sliding the 20 × 20 window. i) Final detection Journal of Multimedia Processing and Technologies Volume 4 Number 2 June 2013

51

(a)

(b)

(c)

(d)

(e)

(f )

(g)

(h)

(i)

(1) (2) (3) (4) Figure 5. Detection results on images from the first database. a) The original image. b) Corner detection. c) Corner detection after elimination of small connected components. d) Edge detection. e) Edge detection after elimination of small connected components. f) SVM detection. g) SVM detection after elimination of small connected components. h) Detection results after sliding the 20 ×2 0 window. i) Final detection

52

Journal of Multimedia Processing and Technologies Volume 4 Number 2 June 2013

4. Conclusions And Future Works In this paper a new detection method was proposed which attempts to combine the detection results of three famous methods, namely SVM, corner and edge detectors. The method is supervised and needs color information of the target and nontarget to train the SVM classifier. In the scenes that we have prior information about the color components of the target, the method could be turned to an automatic sea target detection. One of our future works is to extend the method to other methods like car or people tracking. Also, we will try to use the SVM in a manner which does not have needed training samples from each image, so the method could be performed for target detection automatically. References [1] Shi, C., Xu, K., Peng, J., Ren, L. (2008). Architecture of Vision Enhancement System for Maritime Search and Rescue, In: Proc. 8th International Conference on ITS Telecommunications, ITST, Oct., p.12 – 17. [2] Eldhuset, K. (1996). An automatic ship and ship wake detection system for space borne SAR images in coastal regions, IEEE Transactions on Geoscience and Remote Sensing, 34 (4) 1010–1019. [3] Zhang, Y., Huang, W. G., Zhang, Y. G. (2003). On the space remote sensing of vessels at sea with synthetic aperture radar, Chinese Hydrographic Surveying and Charting, 23 (1) 53–57. [4] Jiang, Q. S., Aitnouri, E. M. (2000). Ship detection in radar sat SAR imagery using PNN-model, Canadian Journal of Remote Sensing, 26 (4) 297–305. [5] Novak, L. M., Halversen, S. D., Owirka, G. J., Hiett, M. (1997). Effects of polarization and resolution on SAR ATR, IEEE Transactions on Aero space and Electronic Systems, 33 (1) 102–115. [6] Yang, W., Sun, H., Xu, X., Xu, G. (2004). Detection of ships and ship wakes in space borne SAR imagery, Geomatics and Information Science of Wuhan University (in Chinese), 29 (8) 682–685. [7] Kuo, J. M., Chen, K. -S. (2003). The application of wavelets correlator for ship wake detection in SAR images, IEEE Transactions on Geoscience and Remote Sensing, 41 (6) 1506–1511. [8] Liao, M., Wang, C., Wang, Y., Jiang, L. (2008). Using SAR Images to Detect Ships From Sea Clutter, IEEE Geoscience and Remote Sensing Letters, 5 (2) 194 - 198, April. [9] Yaman, C., Asari, V. (2007). Long-range target classification in a cluttered environment using multi-sensor image sequences, In: Proc. 3rd International Conference on Recent Advances in Space Technologies (RAST), 14-16 June , p. 304 – 308. [10] Yaslan, Y., Giinsel, B. (2004). Detection of sea targets from thermal images, In: Proc. the IEEE 12th Signal Processing and Communications Applications Conference, 28-30 April , p. 672 – 675. [11] Yang, S., He, S., Lin, H. (2008). Video image targets detection based on the largest Lyapunov exponent, In: Proc. the 9th Int. Conference for Young Computer Scientists (ICYCS), Hunan, 18-21 Nov. p. 2973 - 2977. [12] Hong, Z., Jiang, Q., Guan, H., Weng, F. Measuring overlap-rate in hierarchical cluster merging for image segmentation and ship detection, In: Proc. Fourth Int. Conf. on Fuzzy Systems and Knowledge Discovery (FSKD), Haikou, 24-27 Aug, p. 420-425. [13] Hu, Y., Wu, Y. (2008). Number estimation of small-sized ships in remote sensing image based on cumulative projection curve, In: Proc. Int. Conf. on Audio, Language and Image Processing, (ICALIP), Shanghai, 7-9 July, p. 1522-1526. [14] Lombardo, P., Sciotti, M. (2001).Segmentation-based technique for ship detection in SAR images, IEE Proceedings-Radar, Sonar Navigation, 148 (3) 147-159, June. [15] He, S., Yang, S., Shi, A., Li, T. (2008). A novel image moving sea targets detection method based on the natural measure feature, In: Proc. Int. Symposium on Information Science and Engineering (ISISE), Shanghai, 20-22 Dec., 2, p. 397-400. [16] Mirghasemi, S., Yazdi, Lotfizad, M. A Target-based Color Space for Sea Target Detection, Applied Intelligence, Springer. [17] Mirghasemi, S., Banihashem, E. (2009). Sea target detection based on SVM method using HSV color space, In: Proc. IEEE Student Conference on Research and Development (SCOReD), Nov. p. 555 – 558. [18] Mirghasemi, S., Khayati, A. A Neural Network for Sea Target Detection, In: Proc. of 24th Canadian Conference on Electrical and Computer Engineering (CCECE 2011), Ontario, Canada. Journal of Multimedia Processing and Technologies Volume 4 Number 2 June 2013

53

[19] Yazdi, H. S., Lotfizad, M., Kabir, E., Fathi, M. (2004). A New Corner Detection and Its Application in Vehicle Detection, Iranian Journal of Electrical and Computer Engineering, 2 (2) 51-56. [20] Haralick, Robert, M., Linda G. Shapiro. (1992). Computer and Robot Vision, 1, Addison-Wesley, p. 28-48.

54

Journal of Multimedia Processing and Technologies Volume 4 Number 2 June 2013