Enhancing Red Tide Image Recognition using Semantic Feature and Rotation of Algae Image Angle

Transcription

1 Vol.8, No.3 (214), pp Enhancing Red Tide Image Recognition using Semantic Feature and Rotation of Algae Image Angle Sun Park 1, Myeong Soo Choi 2, Yeonwoo Lee 3* and Seong Ro Lee 4 1 School of Information Communication Engineering, GIST, South Korea 2 Institute Research of Information Science and Engineering, Mokpo National University, South Korea 3 Department of Information Communication Engeering, Mokpo National University, South Korea 4 Department of Information Electronics Engineering, Mokpo National University, South Korea 1 sunpark@nm.gist.ac.kr, 2,3,4 {mschoi, ylee, srlee}@mokpo.ac.kr Abstract Red tide is a temporary natural phenomenon involving harmful algal blooms (HABs) in company with a changing sea color from normal to red or reddish brown, and which has a bad influence on coast environments and sea ecosystems. The HABs have inflicted massive mortality on fin fish and shellfish, damaging the economies of fisheries for almost every year from 199 in South Korea. There has been a lot of study on red tide due to increasing of red tide damage. However, internal study of automatic red tide image classification is not enough. Especially, extraction of matching center of image features for recognizing algae image object is difficult because over 2 species of algae in the world have a different size and features. Besides, the accuracy of algae image recognition of various species is low since previous red tide recognition methods mostly use a few species of red tide harmful algae images for training of classification. In order to resolve the above limitation, this paper proposes the red tide algae image recognition method using rotation of image angle and semantic feature based on NMF (nonnegative matrix factorization). The experimental results demonstrate that the proposed method achieves better performance than other red tide recognition methods. Keywords: red tide algae, image recognition, NMF (nonnegative matrix factorization), rotation of image angle, semantic feature 1. Introduction Red tide is a temporary natural phenomenon involving harmful algal blooms (HABs) in company with a changing sea color from normal to red or reddish brown, and which has a bad influence on coast environments and sea ecosystems. Recent wide spread and persistent HABs give seriously impact public health and the economy of fisheries along coasts of many countries, since it mainly kills fish and results in occasional shellfish poisoning. Research on the prediction of red tide events is important in red tide research, which plays a significant role in mitigating the harm of red tide blooms, decreasing the economic loss caused by HABs, and protecting the health and safety of people [1]. Since the red tide of South Korea has a broad occurrence range and spreads quickly, the direct detection of red tide blooms in real time for mitigation activity is difficult. Seven of the 63 species of harmful red tide algae damage fisheries in coastal areas of South Korea. The ISSN: IJSEIA Copyright c 214 SERSC

2 Vol.8, No.3 (214) HABs have inflicted massive mortality on fin fish and shellfish, damaging the economies of fisheries for almost every year from 1993 to 212. Particularly, there were huge harmful algal blooms of cochlodinium polykrikoides in 1995, resulting in about 95 million dollars in fishery damage [1]. There have been many studies on red tide because the damage to sea farming by red tide blooms of harmful algae occurs every year. The research on red tide events of harmful algae in the coastal area of South Korea has mostly focused chemical properties and investigation of biological causes and is generally used for mitigation activity [2]. However, research on a variety of automatic methods for monitoring and recognizing red tide images of harmful algae on the coast of South Korea is still insufficient [1]. The studies of red tide recognition mostly use classification techniques, such as Bayesian classifiers [3], SVM [3], SVDD (support vector data description) [4], and FCM (fuzzy c- means) clustering [5]. These methods [3-5] are difficult to classify red tide algae images except training algae images. In addition, they do not consider matching point between training images and query image, since the training images have different forms regarding various species of red tide algae (i.e., there are blooming red tides with relation to 63 species of algae in the coastal area of South Korea). In our previous study, we proposed the red tide image recognition method using roundness and entropy [6]. This method can recognize various species of red tide algae image, however the method is affected by image angle. In order to resolve the above limitations of few training algae images and without matching point, this paper proposes a new red tide image recognition method using rotation of image angle based on semantic features of NMF. The proposed method has following advantages. First, the method uses rotation of algae images angle to construct image vector set with respect to preprocess for enhancing the accuracy of image recognition. Second, training of the method uses the semantic features of 63 species of red tide algae image by means of NMF which can recognize various species of red tide algae images. Final, our method uses shape roundness and entropy of red tide images which improves the recognition accuracy of algae images to apply matching point of the images for the images recognition. 2. Related Works Recently, studies of red tide recognition mostly use classification techniques as follows. Jiang et al., proposed the red tide algae classification method using Bayesian classifier and SVM (support vector machine). Their method removes unknown and polluted algae images using Bayesian classifier. Red tide algae images are classified using SVM [3]. However, there is no good performance to remove unknown and polluted images. Besides, they use only 8 species of red tide algae for training images classification. Jiang et al. proposed red tide algae recognition method using two step classifiers. Their method uses first step classifier in which SVDD (support vector data description) removes unknown and polluted algae images. Algae images are classified into the predefined class label using second step SVM classifier. Also this method uses only 8 species of red tide algae for learning image classification [4]. Lili et al., use SVM and FCM (fuzzy c-means) clustering method for classifying red tide images. Their method uses SVM to identify class of test sample. The method identifies the misclassified data using probability based method, and then the misclassified data is reclassified using FCM. Their method also uses 5 species of red tide algae to train image classification [5]. 56 Copyright c 214 SERSC

3 Vol.8, No.3 (214) 3. Non-negative Matrix Factorization This section reviews NMF theory with algorithm and describes the advantage of semantic features of NMF in Example 1. In this paper, we define the matrix notation as follows: Let X *j be j th column vector of matrix X, X i* be i th row vector, and X ij be the element of i th row and j th column. NMF is to decompose a given m n matrix A into a non-negative semantic feature matrix W and a non-negative semantic variable matrix H as shown in Equation (1) [7]. A WH (1) where W is a m r non-negative matrix and H is a r n non-negative matrix. Usually r is chosen to be smaller than m or n, so that the total sizes of W and H are smaller than that of the original matrix A. The objective function is used minimizing the Euclidean distance between each column of A and its approximation which was proposed by Lee and Seung [7]. As an objective function, the Frobenius norm is used: E m n 2 ( W, H ) A WH A ij W F il H lj (2) i 1 j 1 l 1 Updating W and H is kept until converges under the predefined threshold or exceeds the number of repetition. The update rules are as follows: r 2 H T ( W A ) H T, ( W WH ) W T ( AH ) i i W i T (3) ( WHH ) i Example 1) We illustrate an example of NMF and SVD (singular value decomposition) [2, 1]. The non-negative matrix A is decomposed by nnmf() function of Matlab 7.8 into two non-negative matrices, W and H, as shown in Figure 1(a). Figure 1(b) shows the result of SVD by svd() function of Matlab 7.8. There are no zero values and negative values in relation to the semantic feature matrices U and V in Figure 1(b). Unlike SVD, the semantic feature matrices W and H by NMF are sparse in Figure 1(a). Intuitively, the NMF can obtain semantic features that have a small semantic range rather than SVD. In other words, the sparse property of the semantic features of NMF can cover class labels by several pixels of image to be associated with the semantic features. Thus, the semantic feature can easily identify class label pixels to signify image cluster [7]. A W H (a) Result of NMF decomposition of A matrix Copyright c 214 SERSC 57

4 Vol.8, No.3 (214) A U S V Proposed Method (b) Result of SVD decomposition of A matrix Figure 1. Example of NMF and SVD This paper proposes a red tide algae image recognition method using the semantic feature of images and the preprocessing algae image. The proposed method consists of three phases: preprocessing, candidate image extraction, and recognizing red tide algae image, as shown in Figure 2. In the following subsection, we explain each phase in full. Training algae image set Query image (a) preprocessing - rotation of image angle - fixing image size - normalization (b) candidate image extraction - calculating semantic feature of image by NMF - calculating semantic distance between query image and training image (c) recognizing red tide image - calculating roundness of image - computing entropy of image 4.1. Preprocessing Recognized image Figure 2. Proposed Red Tide Image Recognition Method Example 2 shows that the different red tide algae images are recognized by different angle of same query image. In order to solve the above limitations, we rotate the algae image angle. The preprocessing phase of Figure 2(a) consists of three steps: rotation of image angle step, fixing image size step, and normalization step. The phase is described as follows. First, rotation of image angle step rotates algae image angle according to based angel. Second, fixing image size step converts different algae image pixel size into same algae image pixel size (i.e., 12 by 1 pixel matrix). In addition, the center of image is almost reconciled with the center of image vector. Final, the normalization step normalizes all red tide algae image by values of mean and variance of image vectors in order to decrease an error occurring of image background and light. Example 2) Figure 3 shows results of image recognition in connection with changing image angle by using the modified Turk s method [8] based on SVD. We use his method for our basic algae recognition method since Turk s method is fast, relatively simple, and works well in a constrained environment. Figure 3(a) shows recognized image 9.jpg (i.e., eigen-image of 58 Copyright c 214 SERSC

5 Vol.8, No.3 (214) red tide algae of Cochlodinium P.) with regards to the query image q1.jpg. But the query images of 6 and 24 angle in Figure 3(b) recognize eigen-image of Prorocentrum T. (39.jpg). In Figure 3(c), the query image of 18 angel searches eigen-image of Lauderia B. (6.jpg). In Figure 4(d), the query images of 12 and 3 angles recognize eigen-image Rhizosolenia I. (55.jpg). (a) original image (b) 6 and 24 angles recognized image query image q1.jpg 9.jpg (c) 18 angle recognized image query image q3.jpg 39.jpg (d) 12 and 3 angles query image q7.jpg recognized image 6.jpg query image q11.jpg recognized image 55.jpg Figure 3. Results of Image Recognition with Respect to Image Angle Variation of Cochlodinium P 4.2. Candidate Image Extraction The candidate image extraction phase of Figure 2(b) constructs candidate image for training and recognizing of algae image. The phase consists of the calculating semantic feature of image step and the calculating semantic distance step. In the first step, the semantic feature of image is calculated using the NMF algorithm of Equation (1). We modified Turk s method [8] by using NMF for training and candidate algae image set. The training and candidate algae image set are constructed by the modified Turk s method [8] based on the image semantic feature of NMF. In the second step, the candidate images are extracted having minimum value of Euclidean distance between a query image and the trained image set. However, our proposed method has a problem of without matching features between training images and query image, since the training images have different forms regarding 63 species of red tide algae in South Korea Recognizing Red Tide Image Our above NMF based method or algae image recognition methods of related works have problems in recognizing different image according to viewpoint of query images. Because there is no matching point between query image and candidate image set. In order to resolve the problem of without matching point, this paper proposes the roundness and entropy based algae image recognition method to enhance the quality of image recognition. The algae images can be recognized to compare roundness of shape between query image and candidate images (i.e., training image set). However, each 9 groups with relation to species of red tide algae have very similar roundness values. Figure 4 shows sample of red tide algae images with regard to roundness. Copyright c 214 SERSC 59

6 Vol.8, No.3 (214) (a) ~.1 (b).1~.2 (c).2~.3 (d).4~.5 (e).5~.7 (f).7~.8 (g).8~.9 49 (Leptocylindrus D.) 16 (Eutreptiella G.) 15 (Dinophysis T.) 1 (Dictyocha F.) 24 (Gyrodinium S.) 19 (Gymnodinium C.) 3 (Noctiluca S.) Figure 4. Sample Red Tide Algae Images Regarding Roundness Values To solve this limitation of similar roundness values, we use entropy of image to recognize red tide algae image with respect to query algae image. The proposed method in the Figure 2(c) is described as follows. First, the roundness of algae image objects are calculated by Equation (4). Second, a candidate set for algae image recognition is composed of having top 1 roundness value in connection with query image. The entropy of query and candidate images is calculated by Equation (5). Final, the image having highest entropy value is selected to recognize query algae image. In this paper, the roundness scale goes from to 1. (i.e., the roundness of circle shape and horizontal shape is 1 and respectively.). Roundness R() is defined as follows [9]. R ( I ) 4 area perimeter 2 (4) where I is 12 1 image matrix, π is pi that is the ratio of the circumference of a circle to its diameter (i.e., pi is ), area is area of image object, perimeter is a path that surrounds an area of image. Entropy of Equation (5) is given as follows [9, 1]: m Entropy(I ) p log (p ) (5) st i 2 i where I st is 12 1 semantic feature image, p is number of histogram of I st. 5. Experiment This paper uses images of 63 species of red tide algae in the coastal area of South Korea for training of red tide images. The images are obtained from the red tide information on Korea marine environmental information site [11] and Google image search [12]. We used the training data of 3,15 algae images with respect to 63 species of red tide algae. Fifty images are included in the label of one species of red tide algae. We use 35 query images with respect to 7 species of most harmful algae on South Korea. Average accuracy is used to measure the recognition performance. We have conducted performance evaluation by testing proposed method and comparing it with 8 other recognition methods using same training and query images. We implemented 9 image recognition methods: PCA, NB, NN, SVM, B+SVM, SVDD+SVM, SVM+FCM, PCA+RE, and RNMF. RNMF denotes our proposed method in this paper. PCA+RE denote our previous proposed i 1 6 Copyright c 214 SERSC

7 Vol.8, No.3 (214) method [6]. PCA denotes Turk s method using PCA [8]. NB denotes image classification method using naïve Bayesian [1]. NN denotes image classification method using back propagation neural network [1]. SVM denotes image classification method using support vector machine [1]. B+SVM denote Jiang s method using Bayesian and SVM classifiers [3]. SVDD+SVM denote Jiang s method using support vector data description and SVM [4]. SVM+FCM denote Lili s method using SVM and fuzzy c-means [5]. The average accuracy of evaluation results of RNMF is approximately 17% higher than that of PCA, 14.6% higher than that of NB, 16.8% higher than that of NN, 12.99% higher than that of SVM, 1.5% higher than that of B+SVM, 8.9% higher than that of SVDD+SVM, 6.3% higher than that of SVM+FCM, 4.8% higher than that of PCA+RE. 6. Conclusion In this paper, we proposed red tide image recognition method, and the architecture of a system to implement it. The proposed method uses semantic feature image of NMF from 63 species of red tide algae in coastal areas of South Korea for training algae images which can recognize various species of red tide algae images. Besides, the method uses rotation of image angle and image roundness values to apply matching point between query image and trained images. And it uses entropy of red tide images which improves the recognition accuracy of algae images. Our method is simple and works well about various species of red tide algae. Acknowledgements This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (No ) References [1] Red Tide Information System, Korea National Fisheries Research & Development Institute, (212). [2] Y. M. Kim, Y. G. Byun, Y. Huh and K. Y. Yu, Detection of Cochlodinium Polykrikoides Red Tide Using MODIS Level 2 Data in Coastal Waters, Journal of Korean Society of Civil Engineers, (27), pp [3] T. Jiang, C. Wang, B. Wang, J. Xie, N. Jiao and T. Luo, Real-time Red Tide Alge Classification Uing Bayes Classifier and SVM, Proceeding of International Conference on Bioinformatics and Biomedical Engineering (ICBBE), (28), pp [4] T. Jiang, C. Wang, B. Wang, J. Xie, N. Jia and T. Luo, Real-time Red Tide Algae Recognition using SVM and SVDD, In proceeding of International Conference on Intelligent Computing and Intelligent System (ICIS), (21), pp [5] X. Lili, J. Tao, X. Jiozhen and Z. Shaoping, Red Tide Algae Classification Using SVM-SNP and Semisupervised FCM, In proceeding of International Conference on Education Technology and Computer (ICETC), (21), pp [6] S. Park, Y. K. Ji and B. R. Cha, Red Tide Image Recognition using Roundness and Entropy based on Semantic Feature, Advanced Science and Technology Letters (ASTL) 4, (212), pp [7] D. D. Lee and H. S. Seung, Learning the parts of objects by non-negative matrix factorization, Nature, vol. 41, (1999), pp [8] M. Turk and A. Pentland, Eigenfaces for Recognition, Journal of cognitive Neuroscience, vol. 3, no. 1, (1991), pp [9] R. C. Gonzalez, R. E. Woods and S. L. Eddins, Digital Image Processing Using MATLAB, New Jersey, Prentice Hall, (23). [1] J. Han and M. Kamber, Data Mining Concepts and Techniques, Morgan Kaufmann, (26). [11] Red tide information on Korea marine environmental information site, (213). [12] Google Image Search, (213). Copyright c 214 SERSC 61

8 Vol.8, No.3 (214) Authors Sun Park, he is a research professor at school of Information Communication Engineering at Gwangju Institute of Science and Technology (GIST), South Korea. He received the Ph.D degree in Computer & Information Engineering from Inha University, South Korea, in 27, the M.S. degree in Information & Communication Engineering from Hannam University, Korea, in 21, and the B.S. degree in Computer Engineering from Jeonju University, Korea, in Prior to becoming a researcher at GIST, he has worked as a research professor at Mokpo National University, a postdoctoral at Chonbuk National University, and professor in Dept. of Computer Engineering, Honam University, South Korea. His research interests include Data Mining, Information Retrieval, Information Summarization, Convergence IT and Marine, IoT, and Cloud. Myeong Soo Choi, is Research Professor at the Institute of Information Science and Engineering Research at the Mokpo National University, Korea. He received his BS, MS, and Ph.D degree in Electronics Engineering at Mokpo National University, Mokpo, Korea, 2, 22 and 29, respectively. His research interests include WPANs, maritime communication, embedded system, ubiquitous sensor network, and wireless communication. Yeonwoo Lee, he is currentlya Professor with the School of Information Engineering at the Mokpo National University, Mokpo, Korea, since September 25. He has been a Senior Researcher with 4G Mobile Communication team at the Samsung Advanced Institute of Technologies (SAIT), Kiheung, from January 24 to August 25. From October 2 to December 23, he has been a Research Fellow with the School of Electronics and Engineering at the University of Edinburgh, UK. From October 2 to December 22, he joined core 2 work of Mobile VCE program in UK. He received a MS and Ph.D. in Department of Electronics Engineering from Korea University, Seoul Korea, in 1994 and 2, respectively. His research interests are wireless multimedia mobile telecommunication systems, radio resource management, (ad-hoc) multihop relay system, sensor network and particularly their applicable issues to 4G mobile communication systems and cognitive radio systems. 62 Copyright c 214 SERSC

9 Vol.8, No.3 (214) Seong Ro Lee, He received th B.S. degree in electronics engineering from Korea University, Seoul, Korea, in 1987, respectively, and the M.S., and Ph.D. degrees in electrical engineering from Korea Advanced Institute of Science and Technology, Daejeon, Korea, 199 and 1996, respectively. In September 1997, he joined the Division of Electronics Engineering, Mokpo National University, Jeonnam, Korea. His research interests include digital communication system, mobile and satellite communications system, applications of telematics, USN and embedded system. He serves as chairman of detection and estimation committee for the Korea Information and Communications Society. Copyright c 214 SERSC 63

10 Vol.8, No.3 (214) 64 Copyright c 214 SERSC