Performance of RGB, HSI and CIELAB Colorspace in Fuzzy C-Means Classification of Agarwood Image

Transcription

1 Article Performance of RGB, HSI and CIELAB Colorspace in Fuzzy C-Means Classification of Agarwood Image Abstract Colour is one of the quality features used in determining agarwood quality and grade. In this paper, a study on the performance of Red, Green, Blue (RGB), Hue, Saturation, Intensity (HSI) and CIELAB colorspaces in Fuzzy C-Means (FCM) clustering of agarwood 1 image will be presented. The method starts by extracting the agarwood features using an image thresholding technique. These features were then used as the input parameters of 1 FCM algorithm. The performances of each colorspaces were then tested in different number of FCM clusters using five cluster validity indexes named Partition Coefficient 1 (PC), Classification Entropy (CE), Separation (S), Xie and Beni s Index (XB) and Dunn Index (DI) based on the membership degree of each pixels and the value of cluster centre. 1 One hundred and fourty agarwood images in seven different prices were used. Based on the experiment results, CIELAB gave the best performance most of the time compared to the 0 other colorspaces. Keywords: Agarwood, colorspace, fuzzy c-means, classification, cluster validity.

2 Introduction Agarwood is a highly demanded product in perfume industry. Each agarwood product or applications is depending on the agarwood grade. Therefore, grading process is very crucial. Colour of the wood has been used as one of the criteria in determining its quality and hence deciding its price [1-3]. Usually, the higher resin content of wood will give the higher grade and has a shining black in colour. Resin content will change the heartwood colour of agarwood from yellow beige to brown or black. Therefore, colour properties of agarwood can be considered correlates with its resin content [1,,]. In a current practice, agarwood are graded using manual visualization based on the colour appearance. Human characteristics often tender to fatigue where lead to produce misclassification [5,]. There are still no systematic standard to relate agarwood colour and its grade being reported. It is 1 difficult to create a standard because the colour pattern of agarwood is influenced by a lot of parameters i.e. species, defects, illumination, temperature, radiation and chemical substances 1 [7]. Therefore, the grading process is depends on the expert people which had years of experience. 1 The application of image processing and pattern recognition technique has been evolved in automatic grading technology. The unsupervised classification can be used to 1 automatically classified the agarwood colour pixels into several set of colour[]. Fuzzy C- Means (FCM) is one of the popular unsupervised methods used in classification. Each pixel 0 was classified based on its membership degree in the range of 0 to 1. These membership

3 degrees represents the probability of that pixel belongs to a specific cluster. Kuo et. al. [9] has applied FCM to cluster the printed fabrics images in RGB colorspace. Similar approach has been done by Ronghua et. al. c [] for yarn-dyed fabrics in three different colour model i.e. Red, Green, Blue (RGB), Hue, Saturation, Intensity (HSI) and CIELAB. Based on the experiment, the CIELAB colour model gave more effective colour extraction of yarn-dyed fabrics compared to the others. Kang et. al. [11] has applied FCM classification in HSI colour model for dental plaque quantification. Another study done by Saikumar et. al. [1] used FCM algorithm and CIELAB colorspace to find meaningful region of satellite imaginary. On the other hand, several researchers have proposed a combination of FCM with another image processing algorithm or pattern recognition process to find the best region of interest based on colour similarity of a subject [13-15]. As an example, Patmavathi 1 et al. [13] used a combination of FCM and thresholding algorithm to identify underwater images. Grayscale image was used as the input. Chuai-aree et al. [1] has introduced a 1 combination of FCM and neighbourhood smoothing technique in segmenting a document. It gave satisfactory result in defining text, image and background. Ghaleh et al. [15] combined 1 FCM with active contour model in finding a contour of lip. RGB lip colour images were used. The algorithm showed good segmentation of lip even in different speakers with 1 different conditions of illumination. A comparable study of FCM, combination of FCM with morphology and nearest neighbor method has been done by Sopharak et al. [1] to detect 0 exudates in the eyes using HSI colour images. The result showed that FCM gave the highest

4 classification sensitivity (97.%) while a combination of FCM and morphology algorithm gave the highest classification specificity (99.%). Therefore, it can be concluded that the used of FCM in image classification is promising. It can be used to group pixels with similar properties into the same cluster. Cluster validation is very important issue in clustering analysis. The pioneer of the cluster validity was introduced by Pal and Bezdek [17] i.e. Partition Coefficient (PC) and Classification Entropy (CE). Gunderson has proposed Separation coefficient indexes which considered the data geometrical properties. The other indexes were introduced including Fukuyama and Sugeno (FS), Xie and Beni s (XB), Fuzzy Hypervolume (FHV), Partition Density (PD), Separation (S), Separation Compaction (SC), Dunn Index (DI), etc. Most of these indexes measure the compactness and separation in the cluster properties [1-0]. The 1 compactness of cluster measures the variance within cluster to indicates how different each object in the cluster. The separation measures the distance between each cluster (isolation). 1 The good clustering normally will produce high compactness within cluster and high isolation between cluster [19,1]. 1 In this paper, five types of cluster validity indexes i.e. PC, CE, S, XB and DI will be used to measure the performance of RGB, HSI and CIELAB colorspaces in FCM clustering 1 of agarwood image. Two to six number of clusters has been tested by using 10 agarwood images in different colorspaces. The best colorspace was defined as the colorspace which 0 gave the best cluster validity index most of the time.

5 Material One hundred and fourty agarwood images consist of seven different prices i.e. RM50, RM350, RM00, RM900, RM00, RM500 and RM3500 has been used in this study (Figure 1). All of the samples have been graded by an expert at the Malaysian Institute for Nuclear Technology Research (MINT). All algorithms were coded using a Matlab programming language and the program was run on the PC Intel Core Duo. GHz processor with GB of RAM. Methodology In general, the methodology consists of six main steps as follows: 1 Step1: Image acquisition Step : Image segmentation 1 Step3: Colour transformation Step : FCM clustering 1 Step 5: Performance test Step : Decision

6 Figure 1 Seven different samples of agarwood images. The agarwood images were acquired using Nikon D00 CCD camera with the setting of ISO 00, 5mm focal length and without flash. The images were stored in a JPEG format. The camera was setup at fixed distance with 90 angle between the camera and the samples. It was conducted in a control room with fluorescent overhead lighting as the light sources and on the white background colour. Image segmentation is performed to eliminate background and unwanted object in an image. In this study, histogram thresholding was used and applied to the RGB colour images. The process was started by constructing the histogram distribution of each RGB

7 bands for each agarwood price. The threshold value was selected by analyzing the histogram distribution of each agarwood grade. The image has been thresholded as in Eq. 1: where Im(R,G,B) is the original pixel values and TH is the threshold value. Since the (1) agarwood is range from black to brownish colour which is tends to be represented by lower pixel intensity value, therefore, unwanted pixels (background) will be assigned to 55 (white). Only pixels with the value less than or equal to TH will be analyzed during a grading process. In this study, three colorspaces i.e. RGB, HSI and CIELAB has been analyzed. The 1 HSI colour component can be transformed by RGB colorspace coordinate as in Eq. 5. () 1 (3) ()

8 (5) CIELAB colour component is independent and has the ability to measure low colour contrast in an image. The L*, a* and b* component can be obtained by converting the RGB colorspace coordinate into the cube root XYZ matrix transformations (Eq. ) and then the L, a* and b* component is derived as in Eq. 7-9: () (7) () (9) 1 Where, X o, Y o and Z o are the X, Y and Z values with standard white reference.

9 Input Colour Images Band 1 Band Band 3 FCM clustering Cluster center, v Membership Degree, µ No. of cluster = to Max. iteration = 0 = e -5 m= ε Cluster Validity Indexes (PE, CE, S, XB, DI) Colorspace selection Figure Methodology of Colour clustering using FCM algorithm. 1 Figure shows flow chart of colour clustering in this study. Colour images were used as the input in the FCM clustering algorithm. Each pixel is classified based on its 1 membership degree in the range of 0 to 1. These membership degrees will represents the probability of that pixel belongs to a specific cluster. Let say, data X = (x 1, x, x 3,. x n ) 1 denotes an image with n pixels to be classified into c cluster where x i represents multispectral data. The algorithm for an iterative optimization that minimizes the cost 1 function is defined as in Eq. :

10 () where µ ij represents the membership of pixel x j in the ith number of cluster, v i is the ith cluster center and m is a weighing exponent, 1 m. The used of the cost function is to assign pixels with high membership degree with the closer ith cluster center and assign pixels with low membership degree to the far ith cluster center. In FCM methods, the membership degree is depends on the distance between the pixel and each cluster center. The membership degree and cluster centers are defined as in Eq respectively: (11) 1 (1) 1 In FCM algorithm, it starts with an initial guess of the cluster center. The determination of cluster centers is done by iteratively minimizing the sum squared distances

11 between the guess cluster center and the calculation cluster center. This will move all the cluster center to the correct location. The algorithm will stop when one of the number of iteration or termination condition is met. In this study, two to six numbers of clusters has been tested and analyzed. The weighing exponent, number of iteration and the termination tolerance (ε) has been setup to, 0 and e -5 respectively. After clustering process, all of the pixels will have their own membership degree,. Meanwhile, each image will have several cluster centers, v based on number of cluster applied. These data will be used as the input in cluster validity index determination. Since the focus of the colour classification is to find colour set of agarwood based colour similarity, therefore, the main criteria of the selected indexes were based on the variance determination within cluster and distance measurement between each cluster. Five cluster 1 validity indexes i.e. PC, CE, S, XB and DI which met this criterion have been selected. PC Index measures the amount of overlapping between clusters. Where 1/c PC(c) 1 1. The max c n-1 PC(c) produce a best clustering performance for the corresponding datasets [19,]. The index is defined as in Eq. 13: 1 1 (13)

12 CE index measures the fuzziness of the cluster partition where 0 CE(c) log c. The min c n-1 CE(c) produces the best clustering performance for the corresponding datasets [19,]. The index is defined as in Eq. 1: S Index is the index which uses a minimum-distance separation to produce the best (1) clustering performance. The index is defined as in Eq. 15: 1 1 XB index aims to quantify the ratio of the total variation within clusters and the (15) separation of clusters. The best clustering performance is the minimum value of the index 1 []. The index is defines as in Eq. 1: 1 0 Dunn's Index (DI) is originally proposed to be used at the identification of compact and well separated clusters [0,3,]. Where d(c i, c j ) is the minimum intercluster distance (1)

13 from cluster c i to c j, and diam(c k ) is the maximum intracluster distance from one point in cluster c k to another. The maximise DI value is taken as the optimal number of the clusters thus gives the best cluster performance. It can be defined as in Eq. 17. The main drawback of DI are it is computationally expansive and time consuming as c and N increased [0]. (17) Result and Discussion Figure 3 shows a sample of histogram distribution for red band of RM 50 agarwood sample. In general, the pixels are normally distributed in a bell shape. From the observation, a threshold value of provides better representation for extraction of 1 agarwood image in all bands. Therefore, only pixels with the value less than and equal of in all the RGB bands will be used in the analysis. The HSI and CIELAB images were 1 then produced by considering all of these set of pixels.

14 Figure 3 Example of red band histogram taken from RM 50 sample of agarwood. For the performance test, 0 samples of each agarwood grades has been classified into two to six numbers of clusters by using FCM classification. The performance of colorspace is determined based on the maximum frequency of optimal cluster validity of each indexes used. Table 1-5 shows frequency of occurrences for each cluster validity indexes in RGB, HSI and CIELAB in different number of cluster. Table 1 Results of Partition Coefficient Index (PC). Colorspace No. of Cluster, c 3 5 RGB HSI CIELAB Total Table Results of Classification Entropy Index (CE). Colorspace No. of Cluster, c 3 5 RGB HSI CIELAB

15 Total Table 3 Results of Separation Index (S). Colorspace No. of Cluster, c 3 5 RGB HSI CIELAB Total Table Results of Xie and Beni s Index (XB). Colorspace No. of Cluster, c 3 5 RGB 3 1 HSI 19 0 CIELAB Total Table 5 Results of Dunn Index (DI). Colorspace No. of Cluster, c 3 5 RGB HSI CIELAB Total Results from Table 1 have shown that, for PC index, RGB colorspace give the best performance at numbers of clusters. It is followed by HSI colorspace with the value of frequency of occurrence of 5. As the number of cluster increased from 3 to, the PC 1 index was optimal at HSI colorspace. Table shows result of the cluster validity using CE index. It shows the same pattern as the result of PC index. For S index (Table 3), at 1 numbers of clusters, HSI colorspace gives the best performance. However, when the

16 number of cluster increases from 3 to, CIELAB produce the best performance. For XB and DI indexes (Table and 5), the result gave the highest number of occurrences (optimal) at CIELAB colorspace for all number of cluster used. In overall, CIELAB colorspace perform the highest frequency of occurrence of optimal value except for PE and CE indexes. The PE and CE indexes are tends to monotonically decrease with the increasing number of cluster. The consistent performance of CIELAB in S, XB and DI indexes give a strong justification in which CIELAB colorspace provide high accuracy in colour classification by FCM algorithm for agarwood images. The drawback of this method is on the complexity to automatically determine the number of cluster that can provide a significant different colours features and time consuming if high number of cluster is applied Conclusion In this study, three different colorspaces i.e. RGB, HSI and CIELAB had been used to classify agarwood colour using FCM clustering algorithm. The performance of these 0 colorspaces has been determined and compared by applying five types of cluster validity indexes i.e. PE, CE, S, XB and DI. The results show that CIELAB colorspace produce the highest frequency of optimal occurrence for most of the indexes used. This shows that the classification of agarwood colour by CIELAB colorspace proven more accurate results

17 compared to the RGB and HSI. Thus, it can be concluded that CIELAB colorspace provide the best colorspace selection for colour classification of agarwood images using FCM algorithm. Acknowledgments The authors wish to thank and are grateful for the samples and expertise given by Malaysian Institute for Nuclear Technology Research (MINT) References [1] Barden A, Anak NA, Mulliken T, Song M. Heart of the Matter: Agarwood Use and Trade and CITES Implementation for Aquilaria Malaccensis. In: TRAFFIC International Publication; 000 [] Chua LSL. Agarwood (Aquilaria Malaccensis) in Malaysia. In: Forest Research Institute Malaysia; 00 [3] Gunn B, Stevens P, Singadan M, Sunari L, Chatterton P. Eaglewood in Papua New Guinea. In, First International Agarwood Conference. Vietnam: Resource Management in Asia-Pacific Program; 003 [] Asia-Taipei TE, Asia TS. The Trade and Use of Agarwood in Taiwan, Province of China. In: CITES; 005 [5] Abdullah A, Ismail NKN, Kadir TAA, et al. Agar Wood Grade Determination System Using Image Processing Technique. In, Electrical Engineering and Informatics. Institut Teknologi Bandung, Indonesia; [] Donovan DG, Puri RK. Learning from Traditional Knowledge of Non-timber Forest Products: Penan Benalui and the Autecology of Aquilaria in Indonesian Borneo. Ecology and Society 00.9 [7] Kurdthongmee W. Colour classification of rubberwood boards for fingerjoint manufacturing using a SOM neural network and image processing. Computers and Electronics in Agriculture 00.:5-9. [] Chahir Y, Elmoataz A. Skin-color detection using fuzzy clustering. In, International Symposium on Control, Communications, and Signal Processing. Marrakech, Morocco; 00 [9] Kuo C-FJ, Shih C-Y, Kao C-Y, Lee J-Y. Color and Pattern Analysis of Printed Fabric by an Unsupervised Clustering Method. Textile Research Journal :9-1. [] Ronghua Z, Hongwu C, Xiaoting Z, Ruru P, Jihong L. Unsupervised color classification for yarn-dyed fabric based on FCM algorithm. In, Proceedings - International Conference on Artificial Intelligence and Computational Intelligence, AICI [11] Kang JY, Min LQ, Luan QX, Li X, Liu JZ. Dental plaque quantification using FCM-based classification in HSI color space. In, Proceedings of the 007 International Conference on Wavelet Analysis and Pattern Recognition, ICWAPR '07;

18 [1] Saikumar T, Yugander P, Sreenivasa P, Smitha B. Colour Based Image Segmentation Using Fuzzy C- Means Clustering. In, International Conference on Computer and Software Modeling. Singapore: IACSIT, Press; [13] Padmavathi G, Muthukumar. Image segmentation using fuzzy c means clustering method with thresholding for underwater images. International Journal Advanced Networking and Applications 0.: [1] Chuai-aree S, Lursinsap C, Sophatsathit P, Siripant S. Fuzzy C-Means: A Statistical Feature Classification of Text and Image Segmentation Method. International Journal of Uncertainty, Fuzziness and Knowledge 001.9:1-71. [15] Ghaleh VEC, Behrad A. Lip contour extraction using RGB color space and fuzzy c-means clustering. In, 0 IEEE 9th International Conference on Cybernetic Intelligent Systems, CIS 0 [1] Sopharak A, Uyyanonvara B, Barman S. Automatic Exudate Detection from Non-Dilated Diabetic Retinopathy Retinal Images Using Fuzzy C-means Clustering. Sensors 009.9:1-11. [17] Pal NR, Bezdek JC. On cluster validity for the fuzzy c-means model. Fuzzy Systems, IEEE Transactions on : [1] Wu K-L, Yang M-S, Hsieh J-N. Robust cluster validity indexes. Pattern Recognition 009.: [19] Zhang Y, Wang W, Zhang X, Li Y. A cluster validity index for fuzzy clustering. Information Sciences 00.17: [0] Yen C-Y, Cios KJ. Image recognition system based on novel measures of image similarity and cluster validity. Neurocomputing 00.7:01-1. [1] Wang W, Zhang Y. On fuzzy cluster validity indices. Fuzzy Sets and Systems : [] Wu K-L, Yang M-S. A cluster validity index for fuzzy clustering. Pattern Recognition Letters 005.: [3] Chou CH, Su MC, Lai E. A new cluster validity measure and its application to image compression. Pattern Analysis & Applications 00.7:05-0. [] Thilagamani S, Shanthi N. Literature Survey on Enhancing Cluster Quality. international Journal on Computer Science and Engineering 0.: