Intelligent Robotic Inspection System Using Image Processing Technique

Home , Image analysis

INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 8, ISSUE 11, NOVEMBER 2019 ISSN 2277-8616 Intelligent Robotic Inspection System Using Image Processing Technique

Santosh Kumar Sahoo, B.B.Choudhury

Abstract- The objective of the present research work is to develop an intelligent robotic system to identify and classify the defective objects in a bottle manufacturing unit. At this point Artificial Intelligence procedure followed by the different classifier are used in order to classify the defective bottle. The response of the proposed model highlighted that the Least Square Support Vector machine (LSSVM), Linear Kernel and radial basis function has maximum overall performance in terms of Classification rate comparison to others. Thus, the proposed model is a best suitable for classification in a manufacturing unit. The addition of robotic unit also enhances the performance rate as well as more productivity. Here four various machine learning classifiers like KTH NN, ANN, SVM and LSSVM for grouping of defective bottle images. Hence it is perceived that the proposed LSSVM with RBF and kernel are confirmed the higher rate of 96.35 % matched to other classifiers used during the analysis.

Index Terms: Artificial Intelligence. Least Square Support Vector machine (LSSVM), Robotic system, Image Acquisition, Classification rate ————————————————————

1. INTRODUCTION [11] explained the results of some compression of wavelet It is necessary to scrutinize the bottle perfectly in a transforms and thresholding methods. The Support Vector manufacturing unit in order to increase the quality of the Machine (SVM) technique is an intelligent machine learning product. Hence the detection and classification of damaged algorithm having significant performance and more benefits bottle at the manufacturing end is a complicated task. So, over other approaches for solving recent signal and image the proposed method will be much soothing and is suitable processing tasks. Similarly, an ANN specimen is widely for classification of damaged one if trained well, and used in various fields related to non -linear utility estimation subsequently the corrective action will be continuing at a and classification because of its pliability and practicality. At primary phase of the manufacturing process as a result the this stage, our proposed research work aims to develop a quality of the product can be maintained in a smooth system to classify damaged bottle by means of AI- centered manner. classifier which will direct several structures of images and Deepa et. al [1] described an automated vision-based categorize them into a specific group of bottles like defect inspection and sorting system for a plastic injection damaged bottle or defect free bottles. In the meantime, the mold known as a retractor retaining bush by using various identified defective bottle can be segregate by Robot types of sensors and actuators. Xiaoguang et.al.[2] adopt attached with this system. The detailed functional diagram the concepts of multi-instance classifiers for the recognition of an anticipated model is presented in Fig.-1. Here several of mines on the sea- bed from multiple side scan sonar pattern recognition techniques are applied using various views. Jinglong et al. [3] reviews the developments of features pull out from 95 bottle images. Here four various TH wavelet transform (WT) and the applications in rotating machine learning classifiers like K NN, ANN, SVM and machinery fault diagnosis (RMFD). Renjini et al. [4] LSSVM for grouping of defective bottle images. Hence it is described the basic concepts of wavelet transforms and perceived that the proposed LSSVM with RBF and kernel brief idea of recent published works dealing with are confirmed the higher rate of 96.35 % matched to other applications of wavelet theories. Yubin et al. [5] introduced classifiers used during the analysis. a construction of dual wavelet for Image Edge Detection where they used Sobel and Canny filter for comparison purpose. Bruylants et.al. [6] investigated regarding optimally Vision System compression of volumetric medical Images with JP3D and Reference the volumetric wavelets and the entropy-coding improves the Compression performance. Nie et al. [7] explained the image segmentation by histogram thresholding and Class Variance Criterion. Bo Cai1et al. [8] described the image Image sensor Image processor segmentation by using gradient guided active contours. NI-1722 Nitin Sharma et al. [9] described the scheming of complex wavelet transform with directional properties and its Controller application. Tsung-Ching et al. [10] described the variety of arrhythmia Electro Cardio graphic signal utilized for Object to be optimizing scheme of quantization with wavelet techniques inspected for ECG data using Genetic Algorithm (GA). Alarcon et al. ———————————————— Aristo 6 axis Robot  Dr. Santosh Kumar Sahoo is currently working as Associate Professor in the department of EIE, CVR College of Engineering, Hyderabad. India, E-mail: [email protected]  Dr. B. B. Choudhury is currently working as Associate Professor and Fig-1 Schematic functional block diagram of an intelligent Head of Production Engineering at IGIT, Sarang, Odisha, India. inspection system Email: [email protected]

2 METHODOLOGY 2.2. Feature Extraction: Different statistical parameters are calculated using wavelet transform. Initially, four The methodology of the proposed model is sequenced as level decomposition is done, and some features are follows. extracted. The coefficients for them are calculated and using those coefficients, the features (a) Mean, (b) RMS, (c)Standard deviation, (d) Skew-ness, (e) 2.1 Image Acquisition Entropy and (f)power are obtained. This section describes the acquisition of image by an image (a) Mean (  ) : It is usually used to describe numerical sensor like Smart Camera with a 1280x1024 resolution of a bottle from the manufacturing unit. In order to develop a data that is normally distributed. Mean is average of all values. It is referred to a central value of discrete sets datasets different types of bottle images are captured. Block of values. If the values are considered to be numbers diagram for an intelligent classifier model explained in Fig .2. After capturing all the images for different types of bottle, they 1 1 like X1, X2….Xn, then [11]  x   X n are filtered using a band pass filter which removes unwanted M M noise if present. (1)

(b) Standard Deviation (  ): It is a measure of the spread Smoothing by Low Chief Histogram, x Pass filter, Tamura, wavelet of a set of values from the average value. Or in other Gaussian Transforms word it represents spread of data around mean. If there is large standard deviation, then the data is farther from

mean. So small standard deviation means the data is closer to mean. If ―X‖ is a variable with mean  , the Image x Noise Removal Feature standard deviation is given as [11] Acquisition Extraction  1  N  x      X n   x  N  Z n 1   Feature (2) (c) Skewness: It represents lack of symmetry. It is a third Result Classification moment of distribution of data (d) Entropy: It is a measure of randomness of a signal. It is a statistical feature which can be used in representation of an image. Higher the entropy, higher disorder and vice versa. It can be represented as [11] Classifying the KNN, ANN, SVM, n Object LSSVM 2 : e     log( x ) (3) 1

(e) Power: Estimation of power is one of the most Fig.2 Block diagram for an intelligent classifier model frequently used methods of image analysis which provides information about the basic rhythms present After acquiring the bottle image, it is necessary to segment in the image signal and can be easily calculated by it from the image background by an application of image means of Fast Fourier transform i.e., FFT. It can be segmentation techniques. This image segmentation is 2 accomplished by means of gradient and morphological X calculated as[11] : Power=  procedures. This Segmented bottle mostly varies as of the L ( X ) related images by image disparity. An image Gradient can distinguish the discrepancy shown in fig.4 which is again (4) modified to a certain level to acquire segmented images. (f) RMS Value: Which is known as quadratic mean. This The segmented image for defective is displayed in Fig.3. is a statistical measure of the magnitude of a varying quantity which is useful when variants are positive and negative. In case of a set of values, x x x …..x then 1, 2, 3 n the RMS value is given

1 2 2 2 2 by[11]: X  ( X  X  X  ......  X ) . RMS 1 2 3 n n

  1 2  1

  0 0 0 (5)  

 1 2 1  Fig.3 segmented image Fig.4The mask Gradient value

It can be calculated for discrete values. 1 l l 2 coefficient is stated as: e    (w ) Apart from this the various features from an image can cl c, l, I, J collect and reducing the dimensionality. Here the different lxl I 0 J  0 features are extracted from both defective and non- (9) defective bottle pictures. The Tamura texture features Where, w is the coefficient of wave let at [I, J] extraction technique is applied to excerpt different features C , l , I , J like crudeness, distinction and directionality of 95bottle location. As 9 features are attained for a single wavelet images. As Wavelet cantered texture features used coefficient thus for 3 level disintegration of a solo image by normally for the fetidness. So here, the 3 level 2D wavelet using wavelet transform, total 81 features are extracted. decomposition methods are used to evaluation of different The proposed technique is continual for all images, i.e. coefficients of an image in Vertical, Horizontal and Diagonal defective and defect free object class and by gathering all way are presented in pyramidal structure in fig.5. the information related to an image then a data set is formed in order to providing the input variable and the Horizontal output. A binary value ‗0‘ is assigned to represent the Approximate detailed output for a non-defective object and ‗1‘ for a defective Value value of Image-3 object. And by using PCA the dimensions of feature image-3 Horizontal Horizontal detailed value detailed value datasheet are reduced. Vertical Diagonal of image-2 of image-1 detailed Detailed

value of value of 3. RESULTS AND DISCUSSIONS: image-3 image-3 K-NN used as classifier for two classes viz., defective and Vertical detailed value Diagonal Detailed Diagonal defect free. From total database some values are selected of image-2 value of image -2 Detailed for classification and accuracy is obtained value of image-1 (TP  TN ) Vertical detailed value of image-1 Accuracy: = X 100 (10) (TP  TN  FP  FN )

Fig.5 Pyramidal structure of a wavelet decomposition From the table I, it is observed that for different values of ―k‟ different accuracies are obtained. As shown in above table, Again, by considering the above features a training and a for four different combinations i.e. when the training testing data sets are developed in which 95 features of a samples are 90% and testing samples are 10%. Similarly, bottle image is reduced to 25 by means of principal other training and testing samples are chosen such as component analysis (PCA). These training and testing data 85_15, 70_30, 65_35. Accuracy is calculated by selecting are taken as input to the classifier. The proposed structure values of ―k‖ for 3, 5, 6,7,8,9 and 10. Maximum accuracy is reflected 55% circumstances as training of the compressed obtained for the minimum value of ―k‖ i.e. k=3, 5 while for data and others as testing. The individual classifier k=7 also gives maximum accuracy with less execution time. enactment is presented in terms of classification rate (CR) Accuracy has been obtained using formula (10) mentioned can be conveyed as: above.

(TP  TN ) TABLE 1 Classification rate (CR) = (6) ACCURACY AND EXECUTION TIME FOR DIFFERENT VALUES (TP  TN  FP  FN )

Ratio of Training-Testing Value of Accuracy Execution time in Where TP- True positive no., TN= True negative no., FP= Samples "k" in % Second False positive no., FN= False negative no. Now, Wavelet 3 100 1.99 constants are found in a pixel is stated as [12]: 5 100 1.94 1 m  1 n  1 6 80 2.11 w (J , M, N)  F(x, y) φ (X, Y) (7) φ   JM, N KNN_90_10 7 100 2.05 mxn x  0 y  0 8 80 2.05 m 1 n 1 9 100 2.07 I 1 I (8) w (J, M, N)  F(x, y) J, M, N (X, Y) , I  {h, v, d} ψ   mxn x  0 y  0 10 60 2.18

Where 3 100 2.02 J I J J 5 100 1.89  J, m, n (x,y)= 2 2  (P x  m, P y  n) and (x,y)= J m , n 6 100 1.92 J 2 2  ( P J x  m , P J y  n ) are the scrambled and KNN_85_15 7 100 1.89 converted base functions. Where the value of P=2. Now an 8 100 2.09 energy is intended laterally in all the ways like, Vertical, 9 100 1.97 Horizontal and Diagonal (Figure.4). As 3 level of 10 75 2.15 decomposition is used so the energy for each wave let KNN_70_30 3 100 2.05 2557 IJSTR©2019 www.ijstr.org INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 8, ISSUE 11, NOVEMBER 2019 ISSN 2277-8616

5 100 1.98

6 87.8 1.96 7 76 1.94 8 76 2.01

9 87.6 1.9

10 50.2 2.03

3 100 1.98

5 100 1.95 6 77.9 1.98

KNN_65_35 7 66.8 1.91 8 77.9 1.97 9 88.8 1.89 10 77.9 2.05

Also, above results show that optimized value for ―k‖ is obtained to be 3, 5 and 7 where k-NN gives maximum Fig.6. Deviations of error w.r.t total iterations in Multi feed accuracy. Further the accuracy for KNN_90_10 combination neural network classifier of training and testing samples is compared with standard results and it is observed to be giving correct accuracy here Hence it is decided that LSSVM with RBF kernel classifier also. In Fig.6 the deviation of Mean Squared Error (MSE) is the best classifier among all used in the present analysis w.r.t the number of iterations relatively specifies that the for the classification of defective object and defect free MSE regularly declined while the number of iterations object. enlarged more than thousand times. The CR is perceived here as 91.78% shown in Table-2. By considering the 4. CONCLUSIONS: classification degree of the analyzed data, results of support From the results of k-NN classification, it is observed that vector machine classifier are estimated. A proposed SVM for k=3, 5 maximum accuracies are obtained. Further for model competent with linear and Kernel RBF functions in- the k=7, maximum accuracy as well as appropriate order to acquire improved trial information Again the trial computational time is obtained. For four different information can test by SVM linear and radial basis kernel configurations of training and testing samples accuracy is classifier which adjusted and trained earlier. Now the calculated. The performance of k-NN classifier is observed valuation of trial information is articulated through confusion which provides computationally good classification. It is based on distance function therefore for different values of matrix (CR), conveyed in Table-2. Hence Table 2 signifies ―k‖ different accuracies are obtained. So, this method to the presentation of LSSVM classifier in terms of classify the emotions with the help of k-NN classifier is classification rate (CR). Moreover Table 2 also proves that efficient enough. By using the morphological way of image the LSSVM by kernel radial basis function and linear kernel analysis, which forecast the defective bottles in the has maximum precision of 96.35% during object manufacturing process in a precise manner. Current classification. research utilizing the pattern recognition is effectively realized to categorize the defective and defect free bottle from refined principal images. The actual contests are TABLE 2 involved for documentation of structures. These structures THE CLASSIFIERS PERFORMANCE TABLE are effectively categorized as defective and defect free TH bottle by several artificial intelligence classifiers like K Classification nearest neighbour, Artificial Neural Network, Support Vector Classifiers TP FP TN FN rate (CR) Machine, Least Square Support Vector machine. The Multi feed Neural 28 1 11 3 91.78% proposed LSSVM functional blocks effectively network demonstrated as a best analytical tool with advanced Linear kernel SVM 31 10 2 0 86.84% classification rate of 96.35% for analysis. RBF kernel SVM 28 0 12 3 94.06%

LSSVM with linear 29 0 12 2 96.35% Kernel 5. REFERENCES:

LSSVM with Radial [1] Deepa R, Usha S, P V Shashi Kumar, 29 0 12 2 96.35% Basis Function Kernel Advanced Materials Manufacturing & Characterization Vol- 4 Issue 1, pages:26-31, 2014 [2] Xiaoguang Wang1, Xuan Liu1 , Nathalie 2 3 Japkowicz , Stan Matwin ,Journal of Artificial Intelligence and Soft Computing Research.

[3] Jinglong Chena, Zipping Lia, Jun Pana, Gaige Chena, Yanyang Zia, , , Jing Yuanb, Binqiang Chenc, Zhengjia Hea , Volumes 70–71,Pages 1– 35,Elsevier,2015 [4] Renjini L, Jyothi R L, International Journal of Computer Trends and Technology (IJCTT), V- 21(3), Pages: 134-140, 2015. [5] Yubin Li, IJSP, Image Processing and Pattern Recognition Vol.8, No.7, Pages.151-160, 2015. [6] Tim Bruylantsa,,Adrian Munteanua,, Elsevier, Volume 31, Pages.112–133 ,2015. [7] Fangyan Nie and Pingfeng Zhang IJSP, Image Processing and Pattern Recognition Vol.8, No.7, pages.79-88, 2015. [8] Bo Cai12, Zhigui Liu2, Junbo Wang2, and Yuyu Zhu2, International Journal of Signal Processing, Vol.8, No.7, pages.51-62, 2015. [9] Nitin Sharma, Anupam Agarwal and Pawan Kumar Khyalia, Signal & Image Processing: An International Journal (SIPIJ) Vol.5, No.3, 2014. [10] Tsung-Ching Wu, King-Chu Hung, Je-Hung Liu, Tung-Kuan Liu: , Scientific Research,Pages.746-753,2013. [11] V. Alarcon-Aquino1, J. M. Ramirez-Cortes2, P. Gomez-Gil2, The Open Cybernetics & systemic Journal, Vol.-7, pages.32-38, 2013. [12] Dr. Ashish Panat, Prof. Mrs Anita Patil and Miss. Gayatri Deshmukh, IRAJ conference, 2014. [13] Wen-Jiao Zai , International Conference on Wavelet Analysis and Pattern Recognition,2007. [14] S. Samantaa, A. Mandalb, Thingujam, Jackson Singhc, Procedia Materials Science Vol- 6, Pages. 926 – 930, 2014.