Underwater object identification and recognition with sonar images using soft computing techniques
Full text
(2) 666. INDIAN J. MAR. SCI., VOL. 47, NO. 03, MARCH 2018. It consists of minimum three layers10,11. The neurons in the first layer are called input layer. Similarly, the neurons on the last layer is called output layer. The layers present in between these first and last layers are called hidden layers. Each neuron in one layer is connected to every neuron in the next layer and no connections are present between the neurons present in the same layer12-14. As information is continuously fed forward from one layer to the subsequent layers these networks are called feed forward networks as shown in Fig.2. Bias (b=1). Fig.1. Sonar imaging system (ref: www.tes.com). Soft computing based modeling approaches provide an alternative for identification of the system from the available data. Emerging of soft computing began when Lotfi A. Zadeh was working on a fuzzy logic and soft analysis of data. This gave way to the intelligent systems such as neural network, fuzzy logic, genetic algorithm, etc. All the required mathematical concepts are incorporated into a tool box in MATLAB software. Separate toolboxes are available for the neural and fuzzy logic applications for analyzing, designing, and simulating the entire system. The implementation of the proposed work has been done using MATLAB 16 toolboxes. Neural networks-overview A neural network is a computational model which simulates the structure and functions of biological neural networks. It consists of a group of artificial neurons which are interconnected and processes information and generates the output. The activation function possessed by each neuron determines the output of the neuron4. There are many neural network architectures available. Some of the frequently used architectures are Feed forward neural network, Back propagation networks, Pattern Recognition networks and Self Organizing Map. The links between the nodes have weights associated which are altered to achieve an output corresponds to the target output in supervised learning5,6.On the basis of these criteria, the following two different networks such as feed forward network and pattern recognition network with various activation functions have been taken for analysis7-,9.. X1. W1. W -w2. Target /reference image (t). H=f (w1X1 – w2X2)+b. Y=f (H). Comparator. X2 Update weights 0. Fig.2. Structure of Proposedneural network systems (FFN and Pattern Recognition network). In figure, X1 and X2 are inputs, H= hidden layer output, Y = final output of the network, f = activation function. The input neurons processed with tansigmoid activation function as shown in fig.3.The linear activation function has been used in output neurons which gives the output directly proportional to the input shown in fig.4. The Back propagation learning algorithm is used for training15.. Fig.3. Tangent Sigmoid activation function. Feed forward neural network (FNN) In Feed forward neural network, information is passed in the forward direction only..
(3) 667. ANITHA & MALARKKAN: UNDERWATER OBJECT IDENTIFICATION AND RECOGNITION. Fig.4.Linear activation function. Pattern Recognition Network The pattern network is very similar to feed forward network, excluding the last layer and the weights are updated using differential equation16. It is the process of classifying input data into objects or classes based on known features17,18. The input neurons processed with linear activation function give the output directly proportional to the input. The Cross entropy activation function has been used in output neuron given in equation (1). Cross Entropy = −𝑚𝑒𝑎𝑛 𝑠𝑢𝑚 𝑡.∗ 𝑙𝑜𝑔 𝑦 +. Sugeno fuzzy inference system, the two rules stated are given in equation (2&3). Rule 1: If x is A1 and y is B1 then F1 = s1x + t1y + u1 (2) Rule 2: If x is A2 and y is B2 then F2 = s2x + t2y + u2 (3) Here type-3 fuzzy inference system proposed by Takagi and Sugeno is used. A linear combination of the input variables gives the output of each rule added by a constant term for this inference system23,24. The last output is the weighted average of every rule’s output25-27. The corresponding equivalent ANFIS structure is shown in fig.5. There are five layers. Layer 1 and 4 are adaptive and other layers are fixed. The individual layers of this ANFIS structure are described below:. 1 − 𝑡 .∗ 𝑙𝑜𝑔 1 − 𝑦 (1) Where, t = target output and y = pattern network output. Adaptive neuro-fuzzy inference system (ANN) The fuzzy system with its specialist knowledge stands as a frontend preprocessor for the neural network input and output layers. This hybrid system is called the Adaptive network based fuzzy inference system (ANFIS)19-22. In this work, Adaptive Neuro-Fuzzy Inference Systems are fuzzy Sugeno models located in the structure of adaptive systems to facilitate learning and adaptation. Such framework makes Fuzzy Logic Controller more efficient and less relying on expert knowledge. By this hybrid method, an original fuzzy model along with its input variables are derived with the help of the rules extracted from the input-output data of the scheme that is being modeled. Subsequently, the neural network is used to set the rules of the early fuzzy model to produce the final ANFIS model of the scheme. ANFIS structure and parameters Consider the fuzzy inference system has two inputs and one output. For a first order two rule. Fig.5.ANFIS Structure. Layer 1: Every node i in this layer is adaptive with a node function Oi1= µAi(x). (4). Where, x is the input for the ith node and Ai is the linguistic variable connected with this node function and µAi is the membership function of Ai. Usually, µAi(x) is chosen as, µAi(x) = 1/( 1 + [(x−ci/ ai )2]bi) orµAi(x) = exp {−(x – ci/ ai)2 } (5) Where, x is the input and {ai,bi,ci} is the premise parameter set. Layer 2: Each node in this layer is a fixed node which calculates the firing strength wi of a rule. The output of each node is the product of all the incoming signals to it and is given by,.
(4) 668. INDIAN J. MAR. SCI., VOL. 47, NO. 03, MARCH 2018. Oi2 = wi = µAi(x) × µBi(y), i = 1,2. (6). Layer 3: Every node in this layer is a fixed node. Each ith node calculates the ratio of the ith rule’s firing strength to the sum of firing strengths of all the rules. The output from the ith node is the normalized firing strength given by, Oi3 = w̅i = wi / (w1 + w2), i = 1, 2. (7). Layer 4: Every node in this layer is an adaptive node with a node function given by Oi4 = w̅ifi = w̅i (pix + qiy + ri), i = 1, 2. (8). Where, wi is the output of Layer 3 and {pi,qi,ri} is the resultant parameter set. Layer 5: This layer contains only one stable node that figures out the overall output as the summing up of all incoming signals, i.e.. accuracy obtained is more in case of pattern recognition network than the feed forward neural network. Techniques such as Markov Random Field (MRF), Bayes theory and Cross Correlation are based on pixel classification method. These processes were very slow and complex in computation. But in the proposed method, supervised classification techniques based on Feed forward and pattern recognition neural network using GUI have been implemented. Hence the detection is fast, easy and effective. Algorithm 1. The input layer defines the size and type of data which FNN can process. In this plan, the FNN is used to process images, which are 256x256 gray scale images. 2. The hidden layer consists of a single neuron given by the equation H=f (w1X1 – w2X2)+b. Oi5 = overall output, f = ∑ w̅ifi. (9). GUI - GRAPHICAL USER INTERFACE The process of implementing a GUI (graphical user interface) involves two basic steps such as i. Laying out the GUI component ii. Programming the GUI components. In MATLAB, GUIDE is a tool which creates an M-file that contains code to handle the initialization and launching of the GUI. Mainly it needs two files. The FIG-file contains a complete explanation of GUI figure with all controls, grid axis and the values of all object properties. An Mfile contains the functions that launch and control the GUI and the call-backs, which are defined as sub-functions28-31. Materials and Methods Change detection and Object recognition of Sonar image are the two main objectives of this work, which are implemented using a neural network and ANFIS respectively. The accuracy of the respective results was analyzed in the GUI Environment. Change detection The sonar change detection system is designed using feed forward neural network and pattern recognition network. The training network. (10). 3. The output layer Y= f (H), where f= activation function 4. Compare network output Y with the target and adjust the weights. 5. After training the results of the network has been compared with the reference image in order to find the accuracy. The trained network can able to detect changes between the pair of images which are taken at different time intervals. Object recognition Initial fuzzy model can be selected based on the fuzzy rules framed by using the grid partitioning method. Here the input spaces are partitioned into several fuzzy regions to form the antecedents of the fuzzy rules. Grid partitioned fuzzy space for a two-input model, with each input having three trapezoidal membership functions. Two dimensions represent the abscissa and the ordinate of the input space. The rules obtained by this method are then optimized by using ANFIS methodology developed by Jang. This method involves optimization of the premise membership functions by gradient descent algorithm combined with optimization of the consequent equations by linear least squares estimation. The proposed object recognition flowchart has shown in the fig.6..
(5) ANITHA & MALARKKAN: UNDERWATER OBJECT IDENTIFICATION AND RECOGNITION. Input sonar image. Train the ANFIS with input and target pair. Apply image data to ANFIS for extracting the objects. Determine the region properties of the object. Identify the object based on region properties Fig.6. Flow chart for Object recognition. The training of ANFIS with input and target pair is carried out and by applying the 2D gray image to ANFIS the extraction of the objects is done. Here the region properties of the object are determined and the identification of the object based on these region properties is made in the next. 669. The selection of input variables based on regional properties such as Area: The actual number of pixels in the region. Major Axis Length: The length (in pixels) of the major axis of the ellipse that has the same second moments as the region. Minor Axis Length: The length (in pixels) of the minor axis of the ellipse that has the same second moments as the region. Orientation: The angle (in degrees) between the x-axis and the major axis of the ellipse that has the same second-moments as the region. Solidity: The proportion of the pixels in the convex hull that is also in the region. Euler Number: Equal to the number of objects in the region minus the number of holes in those objects. Extent: The proportion of the pixels in the bounding box that is also in the region. step. Thus, with the use of ANFIS, we can detect and identify the required objects in the sonar images of the underwater area. The x and y values are the target values selected from the sonar image by selecting the n- number of places.. Fig.7. ANFIS training network.
(6) 670. INDIAN J. MAR. SCI., VOL. 47, NO. 03, MARCH 2018. Fig.8. ANFIS structure of a trained values. The places selected at white spots is considered as that there is an object and these are considered as the ones, where there are blank spots are considered as no object is present at those places and these are considered as zeros. By these values, a data file is created and these values are loaded in the FIS editor, after loading the data the ANFIS is trained for regional properties. The ANFIS training network and its structure are shown in the fig.7 and fig.8. Results and Discussion The sonar change detection system is designed by supervised clustering algorithms based on neural network. Clustering algorithms used are, feed forward neural network and pattern recognition network. Original scene is used for training the neural network. The changed scene is presented as input and result obtained is subtracted from the original trained scene. A data set used for both the applications like change detection and object recognition consists of a pair of 25 images (taken at different time periods) of different sonar equipment of size 256x256 are used for testing the networks. A pair of image includes an original. image and a changed image of same scene. Out of 25 paired images, 15 are used for training and 10 are used for testing. The results are compared based on accuracy using reference images of same size. The accuracy represents the percentage of correctly determined changed data, i.e., the percentage of the changed data which lies within the reference data. Accuracy varies from 0 to 1. The optimum value of accuracy is 1. The accuracy is computed as given in equation (11). Accuracy in percentage = { 𝑀𝐷𝑃 𝐷𝑃} x 100 (11) Where DP is the number of detected changed pixels and MDP is the number of reference changed pixels matched with detected data pixels. From the results shown in the GUI, it is found that the average accuracy of detection of feed forward network is 80.50% and that of Pattern network is 95.05%. The results can be improved by increasing the number of images in the dataset for training and testing of the network. The GUI shows the output of one of the pair image from the dataset, change detection in Fig.9..
(7) ANITHA & MALARKKAN: UNDERWATER OBJECT IDENTIFICATION AND RECOGNITION. 671. Fig.9. Change detection output through GUI. Fig.10.Object recognition output through GUI. Similarly, a novel algorithm for the detection of underwater objects on sonar imagery is carried out. The sonar object identification system is designed using Adaptive Neuro Fuzzy Inference. System. The proposed algorithms executed by considering blob analysis and region properties. Adaptive Neuro Fuzzy Inference System has been designed for identifying underwater objects and it is.
(8) 672. INDIAN J. MAR. SCI., VOL. 47, NO. 03, MARCH 2018. trained with the images by the spectral values of objects and non-objects. From the analysis, the labeled regions of interest are extracted using region properties. Testing algorithm can extract or identify the objects when similar images are presented. The overall accuracy obtained using equation (11) is 85% in case of object detection. The output GUI for object recognition has given in Fig.10. Conclusion The sonar change detection system is designed using feed forward neural network and pattern recognition network. Accuracy obtained in pattern recognition network is 14.55% more than the feed forward neural network. Similarly, the accuracy of Adaptive Neuro-Fuzzy Inference System for object recognition is 85%. The system has been tested with image pairs of different sonar equipment’s. Proposed system could be extended for images from unmanned underwater vehicles. Improved results can be obtained by implementing high quality images and by increasing the number of database used. Acknowledgements I would like to express gratitude to “GeoMarine Consultants and Technocrats, Chennai” for given me a set of SONAR images for carrying out this research work. References 1.. 2.. 3.. 4.. 5.. 6.. Mignotte.M,. Collet.C, Perez.P, and Bouthem.P, Sonar image segmentation using an unsupervised hierarchical MRF model, IEEE transaction on image processing, vol 9, Issue 7 (2000), 1216–1231. U.Anitha, Study of Object Detection in Sonar Image using Image Segmentation and Edge Detection Methods , Indian Journal of Science and Technology, vol. 9(42), (2016). R. Gonzalez and R. Woods, Digital Image Processing (Addison-Wesley Publishing Company) 1992, 167 168. Mohannad Abuzneid, Ausif Mahmood, Performance improvement for 2-D face recognition using multiclassifier and BPN,IEEE Long Island Systems, Applications and Technology Conference (LISAT),( 2016). Shiranita.K,Hayashi.K,Otsubo.A, Miyajima.T, Takiyama.R,Determination of meat by image processing quality and quality and neural network techniques, Vol.2,(2000), 989-992. Seong-Whan Lee and Hee-Heon Song, A New Recurrent Neural-Network Architecture for Visual Pattern Recognition, IEEE transactions on neural. 7.. 8.. 9.. 10.. 11.. 12.. 13.. 14.. 15.. 16.. 17.. 18.. 19.. 20.. networks, vol. 8, no. 2, (1997). Nielsen.A, The regularized iteratively reweighted MAD method for change detection in multi- and hyperspectral data,IEEE Trans.Image Process.vol. 16, (2007), Issue. 2, 463–478. Scott Reed, Yvan Petillot, and Judith Bell, An Automatic Approach to the Detection and Extraction of Mine Features in Sidescan Sonar, IEEEjournal of oceanic engineering, vol. 28, no. 1, (2003). Ciany.C.M and Zurawski.W, Performance of computer aid edge detection/computer aided classification and data fusion algorithms for automated detection and classification of underwater mines, presented at the Comput.-Aided Detection/Comput.-Aided Classification Conf.,Halifax, Nov. 2001 NS, Canada. Reed.S, Petillot.Y, and Bell.J, An automatic approach to the detection and extraction of mine features in sidescan sonar, IEEE J. Ocean. Eng., vol. 28, (2003) Issue. 1, 90– 105. Wei.S, Leung.H, and Myers.V, An automated change detection approach for mine recognition using sidescan sonar data, in Proc. International Conference. Syst. Man Cybern, (2009), 553–558. Singh, Review article: Digital change detection techniques using remotely-sensed data, International Journal of Remote Sensing, vol. 10, (1989) Issue. 6, 989–1003, Coppin.P, Jonckheere.I, Nackaerts.K, and Muys.B, “Digital change detection methods in ecosystem monitoring: A review, International Journal of Remote Sensing, vol. 25, (2004), Issue. 9,1565–1596,. Bovolo.F and Bruzzone.L, A splite-based approach to unsupervised change detection in large-size multitemporal image: Application to Tsunamidamage assessment,IEEE transactions on geoscience and remote sensing, vol. 45, (2007), Issue. 6,1658–1670. Radke.R.J, Andra.S., Al-Kofahi.O., andRoysam.B, Image change detection algorithms: A systematic survey,IEEE transaction on image processing, vol. 14, ( 2005), Issue. 3, 294–307. Bazi.Y, Bruzzone.L, and Melgani.F, An unsupervised approach based on the generalizedGaussianmodel to automatic change detection in multitemporal SAR images,IEEE transactions on geoscience and remote sensing, Vol. 43, (2005), Issue. 4, 874– 887. Bovolo.F and Bruzzone.L, A detail-preserving scaledriven approach to change detection in multitemporal SAR images, IEEE transactions on geoscience and remote sensing,vol. 43, (2005),Issue. 12, 2963–2972. Shuang Wei and Henry Leung, A Markov Random Field Approach for Sidescan Sonar Change Detection, IEEE journal of oceanic engineering, vol. 37, No. 4, (2012). Tesfaye G-Michael and Rodney G. Roberts, Change detection in sonar images using independent component analysis, Proc. SPIE 10182, Detection and Sensing of Mines, Explosive Objects, and Obscured Targets XXII, 101820K (2017). Anitha.U,Malarkkan.S,Analysis of edge detection techniques for side scan sonar image using block processing and fuzzy logic methods, Advances in Intelligent Systems and Computing(2017)..
(9) ANITHA & MALARKKAN: UNDERWATER OBJECT IDENTIFICATION AND RECOGNITION 21.. 22.. 23.. 24.. 25.. 26.. Isabelle Quidu, Incoherent change detection using amplitude sidescan sonar images, Proceedings of Meetings on Acoustics ,Volume 17, Issue 1,(2012). Subhra kanti Das, Soma Banerjee, Dibyendu Pal, Sambhunath nandy, Sankar NathShome, Somnah Mukherjee, Automatic target detection of sonar images using multi-modal threshold and connected component theory, Indian Journal of Geo-Marine Sciences,(2015), vol.44(2),267-279. Stewart.R. Santos.A, Gustavo.D, Martín del Campo. B., Josué. A. López. R., Intelligent Remote Sensing image post- processing via two-level robust adaptive Neural Network computing, International Conference on Electronics, Communications and Computers (CONIELECOMP),(2014). Sultan Alhusain, Simon Coupland , Robert John , Maria Kavanagh, Design Pattern Recognition by Using Adaptive Neuro Fuzzy Inference System, International Conference on Tools with Artificial Intelligence (ICTAI), (2013) . Jianning Han, Peng Yang and Lu Zhang, Object Recognition System of Sonar Image Based on Multiple Invariant Moments and BP Neural Network, International Journal of Signal Processing, Image Processing and Pattern Recognition, (2014), Vol.7, No.5 ,287-29. Sebastián A. Villar, Gerardo G. Acosta, André L. Sousa and Alejandro Rozenfeld, Evaluation of an Efficient Approach for Target Tracking from Acoustic Imagery. 27.. 28.. 29.. 30.. 31.. 673. for the Perception System of an Autonomous Underwater Vehicle, International Journal of Advanced robotic system, (2014), Vol.11(2). Ayda Elbergui, Isabelle Quidu, Benoit Zerr, Basel Solaiman, Model based classification of mine-like objects in sidescan sonar using the highlight information, Proceedings of meeting on acoustics, (2011). Dario Lodi Rizzini*, Fabjan Kallasi, Fabio Oleari, Stefano Caselli, Investigation of Vision-Based Underwater Object Detection with Multiple Datasets, International Journal of Advanced robotic system, (2015), Vol: 12 (6). Aridgides.T, Frenandez.M, and Dobeck.G, Fusion of adaptive algorithms for the classification of sea mines using high resolution side scan sonar in very shallow water, in Proc. MTE/IEEE OCEANS Conf.Exhib. vol. 1, (2001), 135–142. Reed.S, Petillot.Y, and Bell.J, Automated approach to classification of mine-like objects in sidescan sonar using highlight and shadow information, IEE Proceeding. —Radar Sonar Navigation, vol. 151, (2004), Issue 1, 48–56. Bruzzone.L and Prieto.D. F., An adaptive semiparametric and contextbased approach to unsupervised change detection in multitemporalremotesensing images,IEEE transaction on image processing, vol. 11, (2002), Issue. 4, 452–466..
(10)
Related documents
Drawing up the forecast of the noise mode in the block of cleaning of the shop packagings of vegetable oil and in the territory, adjacent to the room of the block, after
The purpose of this study was explore whether a model of daily health-related utility using EQ-5D derived data from randomised trials would more accurately reflect quality of
Fig. 8 Model for TLR2 immunotherapy ameliorates neurodegeneration in synucleinopathy. In disease condition, TLR2 mediates neurotoxicity. In neuron, i) TLR2 induces
The proposed model named as Cross Data Rate based Aggregation (CDRA) divides packet in MAC queue into groups based on the data rate, packets are to be transmitted at driver domain
Blue indicates variant calls made uniquely by NY-GGC, green indicates variant calls made uniquely by panel testing, yellow indicates common calls made, purple indicates
International Research Journal of Engineering and Technology (IRJET) e ISSN 2395 0056 Volume 02 Issue 03 | June 2015 www irjet net p ISSN 2395 0072 ? 2015, IRJET NET All Rights
In this case, club shares in total Premiership revenue, REVSH%, are estimated as a function of the share of total output measured by playing success in the current and previous
pengaruh dan kesan yang amat signifikan juga dalam menentukan kerjaya