JOURNAL OF MULTIMEDIA, VOL. 9, NO. 3, MARCH 2014
435
Image Retrieval Based on Similarity Learning Zhaohui Liu College of Qinhuangdao, Northeast Petroleum University, Qinhuangdao, Hebei, P. R. China
Rongfu Zhou School of Economics and Management, Yanshan University, Qinhuangdao, Hebei, P. R. China
Abstract—Most traditional image retrieval approaches are based on text surrounding images and fail to capture the content information of the images. These approaches are able to satisfy the present demands. To overcome the limitations of previous approaches, in this paper, we propose a new image retrieval method based on probability similarity learning which is able to exploit probabilistic distribution information. This method involves three stages, image feature extraction, similarity learning and image retrieval. To evaluate the proposed approach, Corel image dataset is used. Corel image dataset contain nine categories which are buildings, birds, flowers, people, trees, elephants, white clouds, mountains and automotive. The experimental results show that the probability similarity learning based approach achieves high accurate in image retrieval network images. For the proposed retrieval method, the accuracy is greatly dependent on the weights. Thus, determining the weights is the next emphasis in the future research. Index Terms—Probability Similarity; Network Image; Image Retrieval
I.
INTRODUCTION
In the recent years, along with the rapid development of computer network techniques, a huge number of digital images appear on the internet. And a large number of pictures are coming into our daily life all the time. So, how to organized such vast image information? As a kind of concrete multi-media information with intuitive expression, images have been paid more and more attention to by people [1]. If the image data could not be managed efficiently, large quantities of information will be lost, and the desired images could not be efficiently retrieved and utilized by people when needed. Therefore, how to help users find their desired images among vast image and manage the images efficiently is becoming a increasingly concerned topic. Presently, the methods for network images retrieval are mainly based on target decomposition, spectral characteristics, fuzzy set and rough set of end member spectral characteristics, Bayesian retrieval method, neural network retrieval method, as well as spectral angle mapping method based on ground material properties and statistical properties of image data, maximum likelihood method and the minimum distance method [2]. And these methods have been widely used in the researches on network image retrieval. However, they face to difficulties in the case of hyper-spectral and multi-angle spectral data. In addition,
© 2014 ACADEMY PUBLISHER doi:10.4304/jmm.9.3.435-441
they have their own shortcomings. For example, for the retrieval method of fuzzy set, its membership is reached depending on the experience or experts, which is highly subjective [3]. So for the learning problems on fuzzy systems, their accuracy and comprehensibility are the primary problem to be solved. Furthermore, when fuzzy clustering technique is used to retrieve network images, the variables need to be selected according to the definition of problem, and the type and characteristics of data in this method. And the rough set is suitable for the retrieval in the case that the data are given. And improper operation may have significant negative influence on the retrieval results [4-5]. When Bayesian retrieval method is used to deal with large-scale image retrieval problems, such a premise as independent “AND” and “CLASS” condition might not be satisfied. Also, it is very difficult to select its desired evaluation function. In addition, the learning procedure is very complicated. In the retrieval, neural network retrieval method has the shortcomings of local minimum and slow convergence, etc.[6]. Besides, the abovementioned traditional retrieval learning method requires high data regularity. Also, they are only used under the premise that the samples are in infinite quantities. However, when network images are retrieved, the data could not always satisfy the above requirements. It usually shows the features of high latitudes, variability and small sample. For these data, it is very difficult for traditional machine learning methods to reach ideal retrieval performance [7]. Besides, the retrieval approach based on end member spectral characteristics is not able to provide the same textural features as that presented by identical permutation. And many network image retrieval methods are based on textural features. For image retrieval based on spectral feature extraction, it could not realize accurate image retrieval. And the traditional retrieval techniques are based on keywords and text, so they could not satisfy the current demands. Along with the appearance of semi-structured and non-structured digital images on the internet, network retrieval technique emerges. In view of the above discussion, a new kind of image retrieval method based on probability similarity learning [8-9] was proposed in the paper. The network image retrieval procedure is composed of three stages, which were preprocessing, image classification and accuracy assessment. (1) In the preprocessing stage, the collected network images were geometrically rectified. And high-
436
JOURNAL OF MULTIMEDIA, VOL. 9, NO. 3, MARCH 2014
Figure 1. The framework of the proposed approach
resolution images were de-noised. Then, the selected images have the same resolution by means of bilinear sampling. (2) In the image retrieval stage, color features were used to represent the images and used for training. In order to avoid such problems as difficult calculation on weighting factor and over-size or under-size image feature in the training process, the training data should be normalized. Before training the extracted image features, grid search technique was used to determine the weighting parameters, which would provide better image retrieval performance and higher retrieval accuracy [10]. (3) Evaluation criterions false negative, true negative, false positive and true positive, etc. were used to quantitatively evaluate the retrieval results based on text. In this paper, the network image retrievals based on rough set and probability similarity learning were compared in the experiment. It was indicated from the experimental results that it was highly accurate to retrieve network images based on probability similarity learning through the experimentation on nine types of samples as buildings, birds, flowers, people, trees, elephants, white clouds, mountains and automotive, etc. in the Corel library and internet. The retrieval accurate rate on network images is highly dependent on the weighting factor. And these parameters significantly influence the retrieval accurate rate and generalization of probability similarity. For the aboveselected probability similarity learning, when calculating, it is more difficult to optimize and determine its parameters. For the traditional method, such optimization methods as genetic algorithm and so on are used to select kernel function. The genetic algorithm has many advantages when solving optimization problems, but it also has many unconquerable disadvantages. Firstly, when dealing with different problems, genetic algorithm needs to design such operators as mutation, selection and crossover, etc once more. Secondly, the operation of genetic algorithm is complicated. In most cases, its calculation is significantly inefficient. In order to overcome the above-mentioned shortcomings of genetic algorithm, grid search method is used to select the weighting factor in the thesis. And it is simple to operate and understand, and has also been widely used in the optimization problems. The advantages of the proposed approach based on probabilistic similarity learning for image retrieval are two folds. (1) To our best knowledge, this is the first
© 2014 ACADEMY PUBLISHER
group of works using probabilistic similarity learning for image retrieval. (2) The proposed approach achieves the competitive performance in comparison with present methods, and can be straightforwardly scaled to other retrieval or recognition tasks. The remaining part of this paper is organized as follows. We propose our similarity learning approach for image retrieval in Section 2 and experimentally evaluate the proposed approach in Section 3. The conclusion is drawn in Section 4. II.
THE PROPOSED SCHEME
To overcome the limitations of previous approaches in network image retrieval, we propose a new approach in this section. The proposed approach for image retrieval is graphically illustrated in the following figure. It is composed of six steps: image collection, image preprocessing, feature extraction, similarity learning, parameter setting and accuracy evaluation. A. Feature Extraction Using Rough Set The information sheet is defined as S U , A,V , F , where
U x1 , x2 ,
A a1 , a2 ,
x10
is
the
research
set;
a3 is the attribute set; V is the range, such
as Va1 weight, medium, light . The information function is expressed as f:U A V . It can also be expressed as follows [1] f x, a Va , a A, x U . Suppose that S U , A,V , F is sub-set [2] P A . The in-distinction relationship must satisfy the following equation: f x, a f y, a , a P . The elements in the basic set are defined as [3] xi Ind ( P ) . Suppose X U , P A , the lower limit is defined as
PX down xi U xi Ind ( P ) X , and the upper limit is
defined as. The critical value is obtained through the following equation PNX PX PX down . (1) If the element X satisfies PX down , PX U , it is considered that the element X is within the data set, which can be seen as the element of the rough set. (2) If the element X satisfies PX down , and PX U , it is considered that the element X is outside of the data set, which can’t be seen as the element of the rough set. (3) If
JOURNAL OF MULTIMEDIA, VOL. 9, NO. 3, MARCH 2014
437
the element X satisfies PX down , and PX U , it is considered that the element X is within the data set, which can’t be seen as the element of the rough set. (4) If the element X satisfies PX down , and PX U , it is considered that the element X is outside of the data set, which can be partially seen as the element of the rough set. The retrieval error rate is calculated through the following equation [4]:
P ( X )
card ( PX down ) , card ( PX )
0 P ( X ) 1
(1)
The attribute simplification aims to eliminate the unnecessary ai to satisfy Ind ( A) Ind ( A ai ) .The variable is defined as X1 , X 2 ,
X n , Xi U , ,
X i X j , X i X j Ui where is a kind of
expressed as equation (3), whose denominator is the ratio of all the users. p u, x p u
||| L(u) | ||| A |
(5)
The following figure is the matrix diagram of the selected commodity. In this figure, the orange points mean the commodities users like, and the unmarked part means the commodities users unlike [17]. Therefore, based on the above definition, the probability of the orange points in the figure can be expressed as the following equation, according to the definition in the paper. A 1 U 1 #P L u 1 L x 1
(6)
retrieval PX down xi U xi Ind ( P ) X 0 mode of the rough set; X i is retrieval form; the upper and lower
limits of are Pdown PX1down , PX 2down ,
, PX ndown
and P PX , PX , , PX . The retrieval accuracy rate of is expressed as follows: n
P ( ) i 1
card ( PX idown ) card (U )
(2)
Suppose P A , and both P () are the same, it should satisfy R P A .For that the following equation is satisfied R () P () .So, RED ( P) could be expressed as the minimum attribute. The core of the minimum attribute set can be expressed as CORE ( P) RED ( P) RED ( P) RED ( P) . B. Probability Similarity Learning Similarity learning has been widely applied to the measurement on cosine and variables, Euclidean distance, Pearson's correlation coefficient, and Jaccard coefficient, etc. Similarity measurement is a learning method based on probability. And its learning theory can be expressed as follows: firstly two random attributes are selected. Meanwhile, the probability of these two attributes will also be selected. The prediction on attribute probability by means of learning algorithm is highly accurate. The probability similarity learning algorithm is defined as: L(y) y | r(u, y) , L(u) {x | r(u, x)}
(3)
where x is the set of the attributes y and u. and among these numerous data, the probability that the record x is selected is expressed as the following equation, whose denominator is the ratio of all the attributes. p u, x p(x)
| Lx | || U ||
(4)
The above equation is calculated from the perspective of data item. And from the perspective of users, it can be
© 2014 ACADEMY PUBLISHER
Figure 2.
Commodity matrix
Thus, by combining the above cases, the probability that users select the commodity x in the thesis can be expressed as the following equation, p u, x
#P #P #N
1
1
A Lu U L x L u Lx
(7)
From the above figures, we could find that: the probability of the set u is the same as that of the commodity x. However, their difference is that: in the left-hand figure, fewer orange points are on the left than right; while it is in the opposite case in the right-hand figure. In the different figures, the calculated probabilities are different either for set u or set x. And in this paper, on the basis of the original algorithm, a weighting factor π is added into the algorithm to eliminate the influence of the probability, which is expressed as, 1 p u, x A L(u) U L(x) 1 π L(u) L(x) V where π can be expressed as π , and v UA V represent the number of all orange points in the figure. In the thesis, the probability that the user u likes the commodity x is defined as follows,
438
JOURNAL OF MULTIMEDIA, VOL. 9, NO. 3, MARCH 2014
p u, x
1 1
V U L(x) UA V L(x)
(8)
d ( R ave)2 (G ave)2 ( B ave)2
C. The Definition of Similarity The probability that users select both the commodities x and y can be expressed as equation (34). e x, y
p(u, x)
(10)
uL(y)
The intersection of the commodities x and y can be expressed as c x, y ||| L(x) L(y) | , where the similarity of the commodity can be expressed as: sim x, y
III.
c(x, y) e(x, y)
average. And the average of d is defined as the difference of grayscale D, whose equation is expressed as follows: 1 N , N is the N i 1 d i number of the pixels. When the three values of RGB are the same, the image is purely gray. Generally pure gray images are in this type. It is found in the equation that: when those three values are almost the same, the image will be nearly colorless. In the contrary, when those three values are different, the image will become colorful.
where ave ( R G B) / 3 , D
(11)
EXPERIMENTS
In this section, we will evaluate the proposed approach for image retrieval. The proposed approach is based on the probabilistic similarity learning. The experimental procedure is illustrated in the following figure 3. The experimental procedure for this approach is composed of five main steps as follows: image preprocessing, feature extraction, probabilistic similarity learning, image retrieval and experimental evaluation. A. Database In the experiment of this thesis, the data comes from the Corel image library and internet. The Corel image library is a popular standard database in the areas of image retrieval and image annotation presently. And 10 types of images and totally 3000 images are collected in the database, which are irrespectively buildings, birds, flowers, people, trees, elephants, white clouds, mountains and automotive, etc. And 1000 samples are collected for each type. The images contained in the database are in a broad range. Among these images, buildings and automotive are man-made objects or artificial scenes; while birds, flowers, people, trees and elephants are natural objects; while white clouds and mountains are typical natural scene. The samples in the database are obtained in the different places, angles and illumination, etc. so the database is highly typical and challenging, which can evaluate the proposed image retrieval method comprehensively and credibly. B. Image Feature Extraction For the image features, color feature should be one of the most important and intuitive perceptive features in the image vision. And the color feature extraction is easier than other image feature extraction. Also, in most cases, satisfied results can be obtained. So, in the image recognition process, color features are used for retrieval, which is greatly concerned. When the color features of an image are extracted, firstly the difference of grayscale should be defined. D is calculated through square root of quadratic sum of the differences between every pigment of RGB and their © 2014 ACADEMY PUBLISHER
Figure 3. The experimental procedure
Then, the color richness of images is expressed as the following equation: 256 1, if C 0 i E other i 0 0， Here, the RGB color space is transformed into HSV, which is consistent with people’s feeling. And three components are used for expression, which are respectively hue H, brightness V and saturation S. The transformation is expressed in another form: T : RGB HSV Finally, HSV space is quantified evenly into 256 colors. H is quantified into 16 classes; while S and V are both transformed into 4 classes. For convenience, the 3dimensional space HSV is expressed with 1-dimensional feature space [17] as follows: Q : HSV C where C is expressed as C Ci | i 0,1, the color in i .
, 255 , Ci is
JOURNAL OF MULTIMEDIA, VOL. 9, NO. 3, MARCH 2014
C. Implementation of Algorithm Through analyzing the experiment, we find that two major factors have influence on the performance of probability similarity: (1) the selection on data item; (2) the selection and calculation of probability in the data item. The data item, the calculation of its probability and other data items have one-to-one relationship, so when retrieving network images, only proper selection of data attributes and the calculation of their probabilities can provide good image retrieval performance. Many kinds of effective and widely-used probability calculation methods are proposed in the previous researches. In the cases of large-scale samples and high-dimension, the definition methods of data probability, attribute probability and similarity in the second paragraph have good retrieval property and have been widely used. In this experiment, the calculation method of this probability similarity is adopted to retrieve network images. The procedures of grid search method can be elaborated as follows: (1) firstly, the possible value range of every parameter is specified, which can satisfy the requirements usually based on experience; (2) secondly, the step of grid search is determined based on experience. Thus, an N-dimensional grid is constructed in the parameterized coordinate. And the parameter value of each in the grid represents a set of weighting factors; (3) K-fold cross-validation method will be used to calculate the retrieval accuracy rate of each group of parameters in the test. Thus, the contour map of retrieval accuracy rate with the parameter variation can be reached. And the optimal kernel parameter is thereby determined; (4)For more accurate search results, fine grid search is carried out on the area with more accurate test after coarse grid search, namely, search area candidates on the contour are selected, and the step is lowered for search [5]. D. Evaluation Criterion for Algorithms In order to evaluate the proposed method in the paper and its related method comprehensively and credibly, common criteria were adopted for evaluation in the experiment. For the retrieval experiment in the thesis, the retrieval accuracy rate was taken as a evaluation criterion: tp tn Accuracy tp tn fp fn tp (true positive) represents true positive; tn where, (true negative) represents true negative; fp (false positive) represents false positive; and fn (false negative) represents false negative. Though the criteria such as recall and so on can also be used to evaluate the retrieval algorithm, the retrieval accuracy is the most widely-used and intuitive evaluation criterion at present. Considering the two types of retrieval problems here, these two types of samples are respectively marked as positive and negative. And then the numerators of the above equation are true positive and true negative; while the denominators are the number of all test samples. The above equation represents the ratio of all the true samples on all the test samples. For multi-class retrieval problem, the definition based on bi-class can be directly extended © 2014 ACADEMY PUBLISHER
439
into that based on multi-class, namely, the ratio of the number of retrieval samples on the number of total test samples in every class. E. Experimental Results and Analysis This experiment uses the probability similarity learning approach to do the network imagereterival. Through the experiment, it can verify the effectiveness of the proposed method in this paper. In the experimental process, for every type images it first preprocessed to extract the image features, and the features were used to represent the samples. The features were sent to the probability similarity learning searcher to do retrieval experiment on network images. Every experiment was repeated for 20 times, and the recognition accuracy of all times was averaged as the final experimental results. In each experiment, 2/3 network images of all the images were randomly selected from the each class of data samples for the training set; while the other 1/3 network images were taken to for mthe test set. The experimental parameters are the attribute value and the weighting factor π that can eliminate the probability influence. And it put the retrieve accuracy rate as the final evaluation standard. In the experimental results of table 1 we can see the probability similarity method can effectively realize the network image retrieval, the retrieval accuracy is over 92%, and the highest can amount to 97%. The reasons for the above results are mainly from the following three aspects: (1). the grid search method can improve the retrieval accuracy of the probability similarity to image retrieval. This is because the grid search method selects probability similarity weighting factor π based on the distribution information of the input data, which can make the probability similarity learning method have better adaptability. (2) No matter in the set u or set x, in different figures the probability is different. And in this paper it joined the weighting factor π to eliminate the influence on probability based on the original probability similarity learning method, which greatly improves the flexibility and effectiveness of this method for network image retrieval. (3) In the experimental process, during the image retrieval it used the color features to capture the image information. In order to avoid the difficulties of calculating the weighting factor π and the image feature too big or too small, it normalized processing the training data during the experiment. In addition, before training the extracted image features, it used the grid search method to determine the weighting parameters so that we can better retrieve image, which can effectively improve the retrieval accuracy rate [18]. In the second experiment, it respectively used the rough set and probability similarity method for the network image retrieval. By comparing the experimental results, it further verifies the effectiveness of the method in image retrieval network. Each experiment repeated 20 rounds. It put the average recognition accuracy of each round as the final results. In each round, it randomly selected two-thirds samples from each type network image as the training set, and the remaining one-third as a test set. The training sample data is 1000, and it took the
440
JOURNAL OF MULTIMEDIA, VOL. 9, NO. 3, MARCH 2014
TABLE I.
RETRIEVAL RESULTS OF NETWORK IMAGES BASED ON PROBABILITY SIMILARITY LEARNING
Test input Buildings samples(1000) Flowers and birds samples(1000) People samples(1000) Trees samples(1000) White cloud samples(1000) Elephants samples(1000) Mountains(1000) automotive samples(1000) TABLE II. Exp. round 1 2 3 4 5
Method Rough sets Probability similarity Rough sets Probability similarity Rough sets Probability similarity Rough sets Probability similarity Rough sets Probability similarity
Grass land 922 0 0 15 10 9 10 0
Traffic land 0 965 10 31 31 10 10 33
Water 25 0 3 939 939 20 939 939
residents 0 25 10 5 2 951 10 2
mix/lost 43 30 13 10 18 10 20 26
Retrieval accuracy rate 92.2% 96.5% 96.4% 93.9% 93.9% 95.1% 93.9% 93.9%
COMPARISON AND EXPERIMENTAL RESULTS OF ROUGH SETS AND PROBABILITY LEARNING Evaluation criterion True positive(tp) True negative(tn) 559 179 638 221 600 284 694 221 598 9 512 232 615 110 555 245 712 189 715 243
former five rounds of results of the experiment. The results are shown in table 2. The experimental parameters are the attribute value and the weighting factor π that can eliminate the probability influence. It used the retrieve correct samples (tp), retrieve correct negative samples (tn), retrieve wrong samples (fp), retrieve false negative samples (fn) and the retrieval accuracy rate as the evaluation criteria. From the experiment results in table 2 we can find that the retrieval accuracy of the method in this paper is concentrated at 90%, and the highest can be 95%. But the retrieval accuracy rate of the rough set method is concentrated in 70%, and in some cases it can be 60%. The reasons for the above experimental results mainly manifested in the following two aspects: because the membership of rough set approach is given according to experience or experts, there is a high subjectivity. So for the learning problems of fuzzy system, the accuracy and comprehensibility are the primary issues to be solved. When using fuzzy clustering to retrieval the network images, it first need to select variables according to the data type, features and the problem definitions, but these are hard to get the reasonable disposition in the experiment process, so it will greatly influence the retrieval result. The probability similarity learning method is mainly to do the geometric correction for the collected network image and denoising processing for high resolution image in the pretreatment stage, and use the double linear sampling to make the selected images have the same resolution. Before training the extract image features, it used the grid search method to determine the weighting factor so that it can better do the image retrieval, and improve the retrieval accuracy. This paper do the experiment using both fixed parameter and the grid search Kernel parameter, which can verify the effectiveness of grid search method in determining the Kernel parameters. It will both repeat 20 © 2014 ACADEMY PUBLISHER
Crops 10 0 964 0 0 0 11 0
Retrieval accuracy rate False positive(fp) 50 30 10 22 10 33 39 90 20 42
False negative(fn) 212 111 106 63 383 223 241 200 79 0
0.738 0.859 0.884 0.915 0.607 0.744 0.747 0.8 0.901 0.958
rounds experiments based on the fixed weighting parameters and the grid search method to determine the weighting parameters, and put the average recognition rate of the 20 times as the evaluation index. The experimental results are as shown in table 3. The experimental parameters are penalty factor C and σ . The average recognition rate of the 5 times experiment is as the evaluation standard. From the experiment results in table 3 we can find that the average recognition rate under the grid search method to determine the parameters is 88.72% and in the fixed Kernel parameter it is 62.96%. So in this paper the grid search method to determine Kernel parameters showed great superiority. TABLE III. Experimental number 1 2 3 4 5 Average recognition rate/%
COMPARISON OF THE RECOGNITION RATE Recognition rate (constant kernel parameter) 0.561 0.723 0.684 0.498 0.682
Recognition rate (kernel parameter by grid search) 0.867 0.924 0.978 0.823 0.844
62.96
88.72
The reasons accounting for the above results are as follows: the grid search method selected the weight factor based on the distribution information of the input data, and the method has higher practical. Besides, the method can map the low dimensional space to high-dimensional feature space. This is helpful to improve the recognition accuracy. IV.
DISCUSSION
A new kind of image retrieval method based on probability similarity learning was proposed in the thesis. The network image retrieval procedure was classified into 3 stages, which were preprocessing, image classification
JOURNAL OF MULTIMEDIA, VOL. 9, NO. 3, MARCH 2014
and accuracy assessment. (1) In the preprocessing stage, the collected network images were geometrically rectified. And high-resolution images were de-noised. Then, the selected images have the same resolution by means of bilinear sampling. (2) In the image retrieval stage, color features were used to represent image information to be trained. In order to avoid such problems as difficult calculation on weighting parameters and over-size or under-size image feature in the training process, the training data should be normalized. Before training the extracted image features, grid search technique was used to determine the weighting factor, which would provide better image retrieval performance and higher retrieval accuracy. (3) Such indices as false negative, true negative, false positive and true positive, etc. were used to quantitatively evaluate the retrieval results based on text. It was indicated from the experimental results that it was highly accurate for retrieving network images based on probability similarity learning through the experiment on such 9 types of samples as buildings, birds, flowers, people, trees, elephants, white clouds, mountains and automotives, etc. in the Corel library and internet. For the retrieval method proposed in the thesis, its accuracy is greatly dependent on the weighting factor. So determining the weighting factors is the next emphasis in the future research. REFERENCE [1] Kusiak, A., “Decomposition in Data Mining: An Industrial Case Study," IEEE Transactions on Electronics Packaging Manufacturing, Vol. 23, No. 4, pp. 345-353, 2009,
© 2014 ACADEMY PUBLISHER
441
[2] Kusiak, A., “Rough Set Theory: A Data Mining Tool for Serniconductor Manufacturing," IEEE Transactions on Electronics Packaging Manufacturing, Vol. 24, No. 1, pp. 44-50, 2011 [3] Lee, 1. -H., Yu, S. -J. and Park S. -c., “Design of Intelligent Data Sampling Methodology Based on Data Mining," IEEE transactions on semiconductor manufacturing, Vol. 17, No. 5, pp. 637-649, 2008 [4] Sushrni, M., Sankar, K. P., “Data Mining in Soft Computing Framework: A Survey," IEEE Transactions on Neural Networks, Vol., No. 1, pp. 3-14, 2009 [5] PingJia, Haitao Li, HuiLin. The remote sensing image classification research based on SVM. Journal ofSurveying&Scientific, Vol. 33, No. 4, pp. 21-23, 2008 [6] Chunyang Wang. The remote sensing image classification research based on SVM. Journal of Silicon Valley, No. 1, pp. 84-85, 2013. [7] Bo Yin, Jingbo Xia, Kai Fu. Network traffic prediction research based on IPSO chaos support vector machine (SVM). Journal of Computer application research, Vol. 29, No. 11, pp. 4293-4299, 2011 [8] Wu Xue, Business Card Recognition System Based on Digital Image Processing, Journal of Multimedia, Vol. 8, No. 2, pp. 137-144, 2013 [9] QianZhang, Reconstruction of Intermediate View based on Depth Map Enhancement, Journal of Multimedia, Vol. 7, No. 6, pp. 415-419, 2012 [10] Jun Zhang, FengXiong, Dan Zhang, Steganalysis for LSB Matching Based on the Dependences Between Neighboring Pixels, Journal of Multimedia, Vol. 7, No. 5, pp. 380-385, 2012