CrossMark

Available online at www.sciencedirect.com

ScienceDirect

Procedía Computer Science 83 (2016) 520 - 528

The 7th International Conference on Ambient Systems, Networks and Technologies

(ANT 2016)

"Efficient Event Driven Sensing in WMSN Using Zernike Moment"

"Manal Al-Sabhan, Adel Soudani" *

""College of Computer and Information Sciences, King Saud University, P.O. Box 51178, Riyadh 11543, Saudi Arabia""

Abstract

Image based sensing in wireless multimedia sensor network (WMSN) is mainly depend on the capability of the deployed scheme to ensure low-power consumption. In depth, the approach of periodic image transmission to the end user, even after compression, will shortly exhaust the energy of the sensors and dramatically reduces the network life time. Thus, detecting an event of interest and extracting the useful information's to decide at source node will avoid flooding the network with unusable data and contributes to extend the whole network life-time. The efficiency of this approach for image-based sensing, in severely resource-constrained sensors, heavily depends on the complexity of the designed sensing scheme.

The main contribution of this paper is to present a low-complexity scheme using Zernike Moment to detect a new object and to recognize the appearance of a specific target before sending a notification to the end user. The paper presents the specification of the proposed scheme and its implementation on wireless multimedia sensors. It addresses the performance's evaluation in terms of time and energy consumption. The results show the high accuracy of the proposed approach to efficiently recognize the target and notify the end user with interesting performance that overcomes the efficiency of other similar sensing approaches proposed in the literature. © 2016 The Authors. Published by ElsevierB.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.Org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of the Conference Program Chairs

Keywords: WMSN, multimedia sensing, object recognition, low-power.

* Corresponding author. Adel Soudani;

E-mail address: asoudani@ksu.edu.sa

1877-0509 © 2016 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license

(http://creativecommons.Org/licenses/by-nc-nd/4.0/).

Peer-review under responsibility of the Conference Program Chairs

doi:10.1016/j.procs.2016.04.248

1. Introduction

Wireless Multimedia Sensor Networks (WMSN) represent an extension of the scalar sensing in Wireless Sensor Networks (WSN) to enable capturing, retrieving, and processing in real time the multimedia information. The design of WMSN is influenced by many factors such as: sensors energy and resources, Quality-of-Service requirements and highly bandwidth demand2.

Tracking and monitoring target based on multimedia streams need a highly data volumes for processing and transmitting which hazard the energy saving. As a result, power conservation is the main concern for designer of these applications since the energy consumption is proportional to the size of the transmitted data.

Obviously, compression techniques has a significant impact on the reduction of data stream. However, these compressions techniques were reported to be complex and not adequate for low-power processing. Another approach is to reduce the power consumption by reduce the size of transmitted data stream in WMSN. This accomplish by determine at sensor node whether the captured image contains phenomena that would be of interest to the end user or not. This approach would not only reduce the power consumption at the source mote, also it promise a highly impact to significantly unloading the network from useless information.

However, the viability of this approach depends on the efficiency of the scheme that is implemented to process the image at the source sensor. Therefore, a big research effort is required to design a low complexity scheme that is appropriates for sensors capability.

Thus, the main contribution of this paper is to present a new scheme for sensing with multimedia capabilities. The proposed image sensing scheme is based on the idea of event detection before data communication using Zernike moment. The novelty of this scheme is to reduce the communication overhead and per-node power consumption while ensuring efficient notification to the end user. The paper focuses on the specification of the new sensing scheme for target detection and recognition. It addresses performances to this scheme for efficient detection and low-energy processing. Comparison with related solutions shows the powerful of the proposed idea.

2. Related Works

In literature there are many contributions that addressed the design of efficient low complexity techniques for object recognition. Yang et al7, classified approaches of shape-based feature's extraction and representation according to their processing methods. Different functions were presented such as: One-dimensional function for shape representation, Polygonal approximation, spatial interrelation feature, Moments, Scale space approaches and Shape transform domains. In this survey, the Moments techniques were described to be very efficient and accurate for object recognition. Different other research contributions attested about the efficiency of using Zernike moment for feature's extraction and for enhancing the accuracy4, 27. Bashkara et al, presented in11 a framework using the Zernike moment to extract feature of Telugu characters in scanned OCR document then he compare it to feature extracted using Hu moments. Also Karbhari et al, 12 have proved Zernike moment efficiency in detection for Marathi language script. An interesting work was presented by H. Marouf et al13. In their approach, they enhanced the face recognition system to find identical twins. The presented method is based on Zernike moment applied with AdaBoost method to detect the face location. In this work, a face recognition scheme based on Zernike moment and Hermite Kernels to cope with facial expressions was developed. The results have shown high ratio of successful identification. X. Yuan et al15, presented an excellent digital image watermarking scheme based on feature extraction and local Zernike moment transformation. They applied their scheme to images decomposed into binary patches. The presented approach have shown interesting results when tested for geometric distortion and rotation. In M. Hitam et al16 and Oluleye et al17, prove the efficiency of Zernike moment as feature extractor method for image retrieval. In their experiment, they have concluded that Zernike moment is fast computed and proves highly capability of object identification.

In the context of wireless multimedia sensor networks Belongi et al. 18, presented a simple and accurate scheme for object matching based on distance between shapes while Vasuhi et al 19, used the Haar wavelet for object feature extraction. In both works, they did not addressed the power consumption. Zuo et al20, presented a distributed two-hop clustered image transmitting scheme. Their method is a trade-off between computation and processing load

that was reflected in enhancing the network lifetime. In 22, the authors used hardware platform for power conservation which need high estimated implementation cost and not considered as scalable solution. In 22 a new scheme based on quad tree decomposition for image compression is presented. The authors suggested it as an efficient solution for low -power solution in WMSN. In 23 an artificial immune system-based image pattern recognition is presented but the associated energy consumption is very high and we think that is not suitable for low-power processing. Despite of their attractive characteristics, the methods based on Zernike moment where not studied for object detection and recognition in wireless multimedia sensor network.

3. General Approach for Image-based Low-power Sensing

In our proposed sensing solution, the multimedia sensors are designed with smart capabilities to enable remote configuration by the end user for target identification. At the set-up of the application, the vector signature representing this target will be broadcasted to the dedicated wireless multimedia sensor through the network. When a new object is detected, the sensor has to decide whether it is the tracked object using a generated feature's descriptor of the object. It compares it with the reference signature that was loaded by the end user at the set-up process. The main challenge in the design of this approach is the specification of an efficient low complexity scheme capable of identifying targets with low-power and low-memory occupancy cost. Wireless multimedia sensors have limited memory; therefore, to ensure optimal multimedia processing, the designer would have to ensure that memory occupancy is kept as low as possible. For low communication overhead, the object descriptor used for identification would have to be represented by a minimum number of bytes since this descriptor has to be broadcasted to all the multimedia sensors. This approach favours scalability for the dynamic run-time changing of target objects.

4. Specification of The Sensing Scheme for Target Recognition

As being expressed in the introduction, the main contribution of this paper is to design and evaluate the efficiency of a new sensing scheme to achieve target recognition and notification to the end user with low energy cost. In depth, the well-known schemes that have been developed for object identification based on image processing are not very appropriated to be implemented in WMSNs 20-24. More specifically, these algorithms would require specific adjustments to enable their use in the context of limited processing bandwidth and limited memory. The novelty of this paper is to design a new scalable scheme that would satisfy the constraints of limited energy and limited resources of WMSNs. The structure of the proposed scheme for object detection and identification is described by the following sequential steps (Fig 1)

Object Detection Object Extraction from Image Object Normalization

End user notification 4- Matching <- Feature Extraction using ZM

Figurel. Structure of sensing scheme based on Zernike Moments

The basic assumption in the specification of this scheme is that sensor will not send data through the network unless an event of interest for the end user has been detected. For that purpose the Multimedia sensor has to be able to detect a new object that might appear in the field of view of the camera. It locally decides if the detected object is of interest to the application and accordingly the network might be involved for data transmission. The different steps that represent the proposed scheme are:

• New object detection: the detection of a new object is based on the approach of background subtraction. It

divides the image into a set of blocks of 8*8 pixels and then the difference between the corresponding blocks containing the new image and the background image is processed at the pixel level. If the whole difference is greater than a certain threshold (Tth), the new object is supposed to be detected. If we note by (/) and Pn-\(j) respectively, the background values for the newly captured image (n) and the previous one. A new object is supposed to be in the

scene if the condition in (1) is satisfied 15.

>W (1)

• Object extraction: the set of blocks representing the object will be isolated to reduce the processed size of useful information's. In fact, this step reduces both the memory occupancy as well as the energy consumption.

• Normalizing: since the Zernike moment has the rotation invariance property only, translation and scaling invariance should be achieved before applying the extraction features set.

• Feature's extraction: The process of extracting the features of the detected object is the main task that determines the performance of the scheme in terms of identification capability. Some metrics have to be considered while designing this task:

o Low complexity. High power consumption is related to the number of arithmetic operations. Hence, the lower is the number of related arithmetic operations, the lower will be the number of associated clock cycles in the motes and consequently the lower consumed energy will be.

o Limited size feature's vector. The feature vector that represents the object descriptor has to be as short as possible. In fact, the memory constraints of the motes and the low bandwidth that is available in the WMSN makes it preferable to retain the shortest feature vector that would allow efficient target identification.

In our approach we have based on the feature's extraction using of Zernike moments. In fact, Zernike moments are invariant for rotation of the object rotation. And based on further processing, they can be also invariant for translation and scale also. In addition, Zernike moment's method has less computational complexity than other methods such as geometric moments.

The application of ZM promises high recognition capability of the target even under different position and orientations. Zernike moment for a continues image f(x,y) that vanish outside of the unit disk, where {f(xi,yjy. 1 <i<M,l<j< N] , can be calculated according to the following equations: 8,9

Znm = ^pj Sx2+y2^1f(.x,y)Vnm(x,y)dxdy = ^Zf f(x,y)Vnm(x,y) (2)

In equation 2, m = 0, 1, 2...i». The m defines the order of the Zernike polynomial while n is either positive or negative, and represents the multiplicity of the phase angels in Zernike moment. Zernike4 introduced a set of complex polynomials which form a complete orthogonal set over the interior of the unit circle. The form of these polynomials is:

Vnm (x, y) = Vnm (p, 0) = Rnm (p) e^8 (3)

Where p is the length of vector from origin to the pixel and theta is the angle formed by this vector and x axis. The Radial Polynomial is defined as:

n M _ y(n-|m|)/2r_lV _(n-s)!_nn-2s (4)

nnmKPJ — ¿js=0 v LJ

Note that Rn>m =Rn ^m and these polynomials satisfy the orthogonality properties.

Zernike moment is rotational invariant but to achieve scale and translation invariance, a normalization step should be considered as a pre step before calculating Zernike moments. Translation invariance achieved by using regular moment to calculate the centroid location of the original image. The two dimensional geometric moment of order (p + q) of a function f(x,y) is defined as:

mP« = £ /" xpyqf(.x,y)dxdy = SiLi ZJU x^y^f{x,y) (5)

Where p, q =1, 2...ro.

So, transforming the image into a new one whose first order regular moment's m01 and m1o are both equal to zero. This can be done using the following equations by mapping the current image to another one where the image is centroid.

, mlO , , m01

x =-ana y =--(6)

m00 J m00 v 7

Scale invariance is achieved by either maximize or minimize the target such that the zeroth order moment m00 is equal to predetermine idle values. Where m00 define the total number of pixels in a given image and (a) is a scale factor computed as following:

a= I— (8)

Where ß is the objective total number of pixels.

Previous equations will move the image target to the origin and descaled to certain factor. To compute the Zernike moments of a given image, the centre of the image is taken as the origin and pixel coordinates are mapped to the range of unit circle. Those pixels falling outside the unit circle are neglected. To map the image into polar coordinate unit disk, we need to compute the polar distance vector p and the angle for any (x,y) pixel as following:

p = Vx2+y2 (9)

d=tan~1- (10)

• Matching with the reference object: We compared the Zernike moment signature of reference target with the obtained vector signature of the object using mean square errors (MSE). If the difference less than 5%, the detected object is considered very similar to the target and we notify the user. Otherwise the detected object is considered as not essential target.

• Notification to the end user: Once the object is detected, the end user is notified. At that level the mote has to process the notification according to the requirements specified by the end user. Thus, this step presents the possibility of achieving a considerable saving in time and power-consumption. More specifically, the notification may involve the transmission of only a few bits, or of the object feature's vector or, alternatively, an extraction of the image of the object.

5. Implementation and Performances Analysis

5.1. Recognition capabilities of the target

The capabilities of the proposed scheme based on Zernike Moment (ZM) to identify a specific target under translation, rotation and scale variances are tested and evaluated using Matlab. Figure 2 shows the test image set composed of 12 images from left to right: Original reference, translation (up, corner and down), rotation by (30, 55, 65, 90 degrees) and scaling by (maximize by 55% and 65% and minimize by 25% and 35%). We use grayscale images of size (64*64 pixels with 8bpp) and (128*128 pixels with 8bpp).

Figure 2. Image set for testing

We have calculated a Zernike moment up to the order 5 that allowed to extract the feature vector set of 12 different images. The Mean-Square-Error was used to analyse the similarity by matching with the reference signature. Figure 3 illustrates the extracted features vector using ZM compared to the reference signature. The results presented in figure 3 (a, b & c) demonstrate the capability of the ZM method to distinguish the similarities between the signatures of the detected object and the reference one.

Figure 3. Extracted features using Zernike Moment for object with Tranalation (a), Rotation(b), and Scaling (c)

A high correlation are shown between the signatures and a considerable invariance for translation, rotation and scaling is proved. This characteristics elect this method to be efficient candidate that can be applied for efficient target recognition.

In table1, we sum up the obtained results of comparison between the feature vectors of the detected object with the reference one. It shows that the calculated MSE was very low, that shows that the proposed scheme was able to recognize the target and presents a high scalability to the image capturing effects.

Table 1. Mean Square Errors of tested images compared to reference one

Image Top Corner Down Rotate 30 Rotate 55 Rotate 65 Rotate 90 Scale 25 Scale 35 Scale 55 Scale 65

MSE 0.0001 0.0001 0.0002 0.0001 0.00001 0.0002 0.0001 0.0004 0.0048 0.0007 0.0045

5.2. Implementation in Tiny-OS platform and evaluation of energy consumption

The performance of the proposed scheme was estimated for wireless sensors networks (MICA2 sensors) using images of size (64 * 64 pixels 8 bpp) and (128 * 128 pixels 8 bpp). The proposed scheme was implemented and tested using AVRORA emulator for MICA2 platform that is based on ATmega128L microcontroller. This tool allows determining the number of clock cycles and the power consumption for the processing of the different tasks of the

implemented scheme.

Table 2 summarizes the consumed time and energy for the tasks of the sensing scheme. Table 2 shows highest power and processing time consumption is related to the new object detection and contour extraction from the background. In fact, the detection of a new object involves processing of all the pixels of the image, whereas the processing of the other part of the scheme requires less data.

Table 2 Evaluation of our scheme on MICA2

Image size 64 * 64 pixels 8 bpp 128* 128 pixels 8 bpp

Measured Attribute Clock cycles Time (s) Energy (mj) Clock cycles Time (s) Energy (mj)

Object Detection and Extraction Target Normalization Feature Extraction using ZM 3514284 492993 7278526 0.47 0.06 0.98 10 1.5 22 3514284 1968961 23465566 1.9 43 0.26 6 3.18 72

Total scheme without Notification 11285803 1.51 33.5 28948811 5.34 121

Table 3.Notification costs in MICA2

Image size 64 * 64 pixels 8 bpp 128* 128 pixels 8 bpp

Measured Attribute Clock cycles Time (s) Energy (mj) Clock cycles Time (s) Energy (mj)

Transmit 1 byte Notification 69569 0.01 0.3 69569 0.01 0.3

Transmit 12 ZM Feature vectors 834828 0.12 3.6 834828 0. 12 3.6

Transmit region of interest 30610360 4.4 132 61220720 8.8 264

For the notification, instead of exhausting the energy consumption by transfer the whole image data, we have considered in this evaluation three type of notification: the simple notification of the detected object (1 byte notification), feature vectors set, or useful extracted blocks from image. Table 3 summarize the performance evaluation for different notification methods. It is clear from the results shown in Table 3 that according to the type of notification we can gain time and energy compared to the whole transmission of the whole image.

When compared to other similar approaches for multimedia sensing, our scheme presents attractive characteristics regarding complexity and power consumption. Table 4, sums up the most relevant reported solutions in the literature and their characteristics compared to what we proposed as an efficient solutions for low-power sensing.

Table 4. Comparison to other schemes presented in literature

Network Type Distributed Quad tree AlS-based Prioritization Object Centroid Presented

Compression20 decomposition23 Solution 24 scheme 22 Extraction 21 & Solution

Histogra25

Processing Distributed Local Distributed Local Local Local Local

Scheme Image Decomposition AI method Image Blocks Useful data Useful Useful data

Approach Compression and Prioritization extraction data extraction

Compression extraction

Implementation software software software software hardware software software

Scalability * - * - - * *

Complexity & High High Energy High High High Low Low

Energy Complexity Consumption Complexity Complexity Implementation Energy Complexity and

and Energy and Energy and Energy Cost Consumpti Energy

Consumption Consumption Consumption on Consumption

6. Conclusion

This paper presents and analyzes a new sensing approach for target recognition and notification in wireless multimedia sensor network (WMSN). The main idea of the scheme was to unload the source mote and, consequently,

the network from heavy data processing and transmission by detecting the event of interest before sending the information to the user. The proposed scheme based on the use of Zernike Moments demonstrated flexibility and scalability for efficient recognition of the target.

The implementation of this scheme on MICA 2 platform showed interesting performances in terms of processing time and energy consumption. The application of Zernike Moments allowed to recognize the target with more accurate results compared to centroid and histogram scheme25. In addition, it outperforms other proposed schemes in the energy consumption.

As future work, we think that investigating the communication overhead in the network will give more clear idea about the efficiency of this solution for deployment in WMSN. We think also that other methods for features extraction and target recognition such Fourier Descriptors can be evaluated the context of WMSN.

References

1. V. Madhuri et al, "A Study on Smart Dust (MOTE) Technology", IJCSET, March 2013, Vol 3, Issue 3, 124-128.

2. I. Akyildiz et al, "Wireless sensor networks: a survey", Computer networks, Elsevier 2015, Vol 38, Issue 4, 393-422.

3. J. Yick et al, "Wireless sensor network survey", Computer Networks: The International Journal of Computer and Telecommunications Networking archive, V 52 Issue 12, August, 2008, 2292-2330.

4. M. Teague, "Image analysis via the general theory of moments", J.Opt. Soc. Amer., vol. 70, no. 8, pp. 920-930, Aug. 1980.

5. M. Hu, "Visual pattern recognition by moment invariants", IRETrans. Inform. Theory, vol. IT-8, pp. 179-187, Feb. 1962.

6. J. Flusser et al, "Moments and Moment Invariants in Pattern Recognition", WILEY, ISBN: 978-0-470-69987,0ctober 2009

7. M. Yang et al, "A survey of shape feature extraction techniques" in Pattern Recognition Techniques, Technology and Applications, chapter 3, pp. 43-90, InTech, 2008.

8. A. Khotanzad et al, "Invariant image recognition by Zernike moments", in Pattern Analysis and Machine Intelligence, IEEE Transactions on , vol.12, no.5, pp.489-497, May 1990

9. S. Liao, Thesis "Image Analysis by Moments", Electrical and Computer Engineering, the University of Manitoba, Canada 1993.

10. F. Tahmasbi et al, "Classification of benign and malignant masses based on Zernike moments", Computers in Biology and Medicine, vol. 41, no. 8, pp. 726-735, 2011

11. P. Bhaskara et al, "Feature Extraction Using Zernike Moments", International Journal of Latest Trends in Engineering and Technology (IJLTET), ISSN: 2278-621X, Vol. 2, Issue 2 March 2013.

12. K. Kale et al, "Zernike Moment Feature Extraction for Handwritten Devanagari (Marathi) Compound Character Recognition", International Journal of Advanced Research in Artificial Intelligence (2014)

13. H. Maroof et al ," Discrimination Between Twins Faces Using Adaptive LRPCA, and COGNITEC" , 12th Iranian Conference on Intelligent Systems, ICIS2014 , Iran, 2014

14. X.Yuan et al, Feature extraction and local Zernike moments based geometric invariant watermarking, Multimedia Tools and Applications, 2013

15. M. Hitam et al, "Content-Based Image Retrieval Using Zernike Moments for Binary and Grayscale Images", Moments and Moment Invariants - Theory and Applications, GCSR Volume 1, pages 275-288

16. B. Honarvar, "Image Reconstruction from a Complete Set of geometric moments", Department of Electrical, Faculty of Engineering, University of Malaya, Lembah Pantai, 50603 Kuala Lumpur, Malaysia Signal Processing, 2014; 98:224-232.

17. B. Babatunde et al , "Zernike Moments and Genetic Algorithm : Tutorial and Application", British Journal of Mathematics & Computer Science 4, pages 2217-2236,2015

18. S. Belongie et al, "Shapematching and object recognition using shape contexts," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 24, no. 4, pages 509 - 522, 2002.

19. S. Vasuhi et al, "Object detection and tracking in secured area with wireless and multimedia sensor network, " in Networked Digital Technologies, vol. 294 of Communications in Computer and Information Science, pp. 356 - 367, Springer, Berlin, Germany, 2012.

20. Z. Zuo et al, "A two-hop clustered image transmission scheme for maximizing network lifetime in wireless multimedia sensor networks," Computer Communications, vol.35, no. 1, pp. 100-108, 2012.

21. D. M. Pham and S. M. Aziz, "Object extraction scheme and protocol for energy efficient image communication over wireless sensor networks," Computer Networks, vol. 57, no. 15, pp. 2949-2960, 2013.

22. K. Irgan et al, "Low-cost prioritization of image blocks in wireless sensor networks for border surveillance," Journal of Network and Computer Applications, vol. 38, no. 1, pp. 54-64, 2014.

23. G. Nikolakopoulos et al, "A dual scheme for compression and restoration of sequentially transmitted images over Wireless

Sensor Networks," Ad Hoc Networks, vol. 11, no. 1, pp. 410-426, 2013.

24. H. Wang, et al., "Artificial immune system based image pattern recognition in energy efficient wireless multimedia sensor networks," in Proceedings of the IEEE Military Communications

25. M. alhilal et al, "Image based object identification for efficient event-driven sensing in wireless multimedia sensor network", Hindawi publishing corporation, Article ID 850869, volume 2015.

26. F. Xet al, a hybrid algorithm of fast and accurate computing Zernike moments. IEEE, 2007.