A New Multi-view Learning Algorithm Based on ICA Feature for Image Retrieval Fan Wang and Qionghai Dai Department of Automation, Tsinghua University, Beijing 100084, China [email protected], [email protected]

Abstract. In content-based image retrieval (CBIR), most techniques involve an important issue of how to efficiently bridge the gap between the high-level concepts and low-level visual features. We propose a novel semi-supervised learning method for image retrieval, which makes full use of both ICA feature and general low-level feature. Our approach can be characterized by the following three aspects: (1) The ICA feature, as proved to be representative of human vision, is adopted as a view to describe human perception; (2) A multi-view learning algorithm is introduced to make the most use of different features and dramatically reduce human relevance feedback needed to achieve a satisfactory result; (3) A new semi-supervised learning algorithm is proposed to adapt to the small sample problem and other special constrains of our multi-view learning algorithm. Our experimental results and comparisons are presented to demonstrate the effectiveness of the proposed approach.

1

Introduction

With the rapid increase of the volume of digital image collections, content-based image retrieval (CBIR) has attracted a lot of research interest in recent years [16]. However, most of the features adopted in the previous approaches are pixel based or extracted by cutting the image into blocks or regions, and further extract feature from the blocks. Therefore, these approaches are mostly concerned with low-level features, such as color, texture, shape, etc., which can not fully represent the human perception. Actually, people do not perceive the images on the level of pixels or blocks, they are always interested in high-level concepts instead of the low-level visual features. As a result, the gap between high-level hidden concepts and low-level visual features has become one of the challenging problems of CBIR systems, due to the rich content but subjective semantics of an image, which can not be fully recognized by computer. Theoretical studies suggest that primary visual cortex (area V1) uses a sparse code to efficiently represent natural scenes, and each neuron appears to carry statistically independent information [20]. Recent researches have shown that, Principal Component Analysis (PCA), and Independent Component Analysis (ICA) of natural static images produce image representation bases resembling the receptive fields of V1 cells [5]. This kind of results, more specifically ICA T.-J. Cham et al. (Eds.): MMM 2007, LNCS 4351, Part I, pp. 450–461, 2007. c Springer-Verlag Berlin Heidelberg 2007 

A New Multi-view Learning Algorithm Based on ICA Feature

451

results, also come from the learning procedure named sparse coding. This coincidence of results was already mathematically justified through the identification of the link between the ICA and Sparse Coding formalisms. The results reported in above-mentioned experiments are well fitted to parametric (Gabor or Wavelets) models which were broadly accepted as approximations for V1 receptive fields [14]. To this day, there have been several approaches that adopt features through ICA to improve the retrieval performance. For example, the paper [9] showed that the PCA and ICA features can be used to construct similarity measures for the image retrieval, and through comparison, the conclusion is made that the ICA basis method outperforms the PCA basis method. In [10], an unsupervised classification algorithm was presented based on an ICA mixture model. This method can learn efficient representation of images of natural scenes, and the learned classes of basis functions yield a better approximation of the underlying distributions of the data. Based on the former research, it is believed that ICA is able to well discover the basis of human vision. We adopt ICA feature in this paper to further approach the human perception. Instead of simply replace the former general visual features with ICA features, a new utilization of ICA features is proposed. While ICA features is some efficient representation of human vision, the low-level features, such as color or texture, carrying abundant statistical information, are the image representation by computer. In other words, ICA features are the representation of images from human’s view, while the low-level features can be regarded as the computer’s view. Since both of the two views are valuable for the retrieval system, a multi-view learning algorithm is necessary to fully utilize these features. A well-know tool to bridge the gap between high-level concepts and low-level features in CBIR is relevance feedback, in which the user has the option of labeling a few of images according to whether they are relevant or not. The labeled images are then given to the CBIR system as complementary queries so that more images relevant to the user query could be retrieved from the database. In recent years, much has been written about this approach from the perspective of machine learning [17], [18], [19], [24]. It is natural that the users will be more willing to see satisfied retrieval results only by once or twice feedback instead of many times of labeling. This limits the amount of available labeled data, and here comes the demand of semi-supervised learning algorithm, which reduce the amount of labeled data required for learning. Multi-view learning algorithms have been studied for several years, and there exist some significant proposals, i.e. Co-Training [3], Co-Testing [12], Co-EM [13], Co-retrieval [21]. However, these methods’ performance drops dramatically if the labeled data is limited, and they do not take enough consideration of the characteristics of the data and the views. In this paper, we propose a new image feature based on ICA expansion, and the distance between ICA features are also defined. We novelly integrate semi-supervised learning method into a multi-view learning framework called

452

F. Wang and Q. Dai

Co-EMT, and ICA features are introduced as one of the views to further improve the retrieval performance. The rest of the paper is organized as follows: Section 2 introduces how to perform ICA expansion and extract ICA features. The multi-view learning algorithm is described in Section3, followed by the proposed semi-supervised algorithm in each single view detailed in Section 4. Section 5 shows the whole scheme of our CBIR system. The experimental results and some discussions of our algorithm are presented in Section 6. Finally this paper is concluded in Section 7.

2

ICA Feature Extraction

ICA is a recently developed statistical technique which often characterizes the data in a natural way. It can be viewed as an extension of standard PCA, where the coefficients of the expansion must be mutually independent (or as independent as possible) instead of being merely uncorrelated. This in turn implies that ICA exploits higher-order statistical structure in data. The goal of ICA is to linearly transform the data such that the transformed variables are as statistically independent from each other as possible [1], [4]. ICA has recently gained attention due to its applications to signal processing problems including speech analysis, image separation and medical signal processing. So far there have been many kinds of algorithms for ICA expansion. However, some may be computationally demanding or have problem of convergence when dealing with data of high dimensionality. In this paper, we choose a fast and computationally simple fixed-point rule of ICA [8] for image feature extraction in consideration of speed. Furthermore, the convergence of this learning rule can be proved theoretically. Here we apply the method to computing ICA bases of images, the detailed steps are discussed as follows. Firstly, the n-dimensional data vectors x (t) were obtained by first taking n1/2 × n1/2 sample subimages from the available image database. Here t is from 1 to N , which is the total number of data samples for x. In the formulation of the ICA, the data vector is assumed to be mixed by unknown sources, that is x (t) = As (t) =

m 

si (t) ai

(1)

i=1 T

here the vector s (t) = [s1 (t) , · · · , sm (t)] contains the m independent components si (t) for the data vector x (t). A = [a1 , · · · , am ] is a n × m matrix, whose columns are called features or basis vectors. The number of independent components m is often fixed in advance. In any case, m ≤ n, and often m = n. Data x is preprocessed to have zero-mean and unit variance. x ← x − E [x]

(2)

 2 x x←x

(3)

A New Multi-view Learning Algorithm Based on ICA Feature

453

The preprocessed vectors were then whitened using standard PCA so that the resulting vectors v (t) had n − 1 components (one of the components becomes insignificant because of the subtracted mean). The PCA whitening matrix is of the form V = D−1/2 E T , where the columns of the matrix E contain the PCA eigenvectors, and the diagonal matrix D has the corresponding eigenvalues as its elements. Standard PCA is used because it can compress the data vectors into an m-dimensional signal subspace and filter out some noise. W is defined as the m × n de-mixing matrix, so that the purpose of the ICA learning is to estimate W in sˆ (t) = W v (t)

(4)

After this, the generalized fixed-point algorithm described in detail in [8] is applied to finding the independent components of the whitened data vectors v (t). In this algorithm, we first initialize the matrix W by the unit matrix I of the same dimension. The update of wi , denoting the i-th column of W , and the orthonormalization are performed as follows:       T T (5) wi∗ (k + 1) = E vg wi (k) v − g  wi (k) v wi (k) wi (k + 1) = wi∗ (k + 1)/wi∗ (k + 1)

(6)

here E{} denotes the mathematical expectation, wi (k) is the value of wi before the k-th update, while wi (k + 1) is the value after it. In practice it is replaced by sample mean computed using a large number of vectors v (t). The function g (u) can be any odd, sufficiently regular nonlinear function, and g  (u) denotes its derivative. In practice, it is often advisable to use g (u) = tanh (u) [6]. The convergence of this method was proved in [7]. From wi we can obtain the estimation for the corresponding basis vector ai of ICA using the formula a ˆi = ED1/2 wi

(7)

that is, the estimation of the mixing matrix is Aˆ = (W V )−1 = ED1/2 W T

(8)

For a new image, we can extract ICA feature from it through mapping it to the basis and getting the coefficients. The image is first sampled by taking n1/2 × n1/2 subimages from it for K times. Then the prewhitened n-dimensional data vectors x (i) , i = 1, · · · , K are obtained. The ICA feature for this image can be calculated as S = W X, where X is composed by the columns x (i).

3

Multi-view Learning Algorithm

Two kinds of image features are utilized in our system: general low-level feature and ICA feature. As mentioned in Section 1, the general low-level feature representation can be regarded as the view of computer when recognizing the image,

454

F. Wang and Q. Dai

while the ICA feature approximates the view of human. That is to say, an image x can be described by these two features in two views. Previous research proved that if there exist two compatible and uncorrelated views for a problem, the target concept can be learned based on a few labeled and many unlabeled examples. We found the two views mentioned above partially satisfies the condition after some statistical test. This is the similar situation in many real world multi-view learning problems. We use a robust multi-view algorithm called Co-EMT [11] which interleaves semi-supervised and active learning, to handle this problem. It has been proved that Co-EMT is robust in harsh conditions when the two views are not completely compatible and uncorrelated. The algorithm Co-EMT includes training step and testing step, which adopt Co-EM and Co-Testing, respectively. Co-EM [13] is a multi-view algorithm that uses the hypothesis learned in one view to probabilistically label the examples in the other view. It can be seen as a probabilistic version of Co-Training [3]. Let V 1 denotes the view of general low-level feature, V 2 the ICA feature. Denote learning algorithms L, which will be talked about later in Section 4. The Co-EM can be described as follows: Firstly, the algorithm trains an initial classifier h1 in the view V 1 based solely on the labeled examples by the learning algorithm L. Then it repeatedly performs the following four-step procedure: (1) use h1 to probabilistically label all unlabeled examples and obtain their labels N ew1 ; (2) in V 2, learn a new maximum a posterior (MAP) hypothesis h2 based on the labels N ew1 learned in the previous step; (3) use h2 to probabilistically label all unlabeled examples again, and get N ew2 ; (4) in V 1, learn a new MAP hypothesis h1 based on the labels N ew2 labeled in the previous step. These steps are repeated for several iterations. At the end, a final hypothesis is created which combines the prediction of the classifiers learned in each view. Since solely depending on the system’s automatic iterations is insufficient for learning, the user’s feedback should be added to input new useful information to the system. Here Co-Testing [12] is introduced as an active learning algorithm, and Co-EM is interleaved with Co-Testing to form the Co-EMT algorithm. After running Co-EM for several iterations on both labeled and unlabeled examples, the two hypotheses in two views have been trained sufficiently. The data points on which the hypotheses on two views disagree the most consist the contention set, which means we are least confident on the label of these samples using the two hypotheses. Labeling these points by the user can provide the system with the most information from the user’s perception, thereby enhance the effectiveness of the learning algorithm.

4

Proposed Semi-supervised Learning Algorithm in Each Single View

In the view of general low-level feature, we use Euclidean distance as the distance measure between any two images xi , xj :

A New Multi-view Learning Algorithm Based on ICA Feature

d(xi , xj ) =

455

xi − xj 2 if xi − xj 2 < ε ∞ otherwise

where ε is a positive threshold to assure the sparsity of the distance matrix. Since the images in positive set R have been labeled relevant, we set the distance between each of them as zero, that is, d (xi , xj ) = 0, ∀xi , xj ∈ R. In the view of ICA feature, we also need distance measurements between each pair of the features. According to the equations in Section 2, we firstly use the labeled positive examples to train the basis vectors which expand the ICA subspace corresponding to the positive set. For an image x, we sample subimages from it, and map the subimages to the acquired basis vectors to obtain the m×K coefficient matrix S, which we treat as the feature of image x. Here K is the number of patches sampled from x. Each column of S is a vector in m-dimensional space, and all the K columns in the feature S of image x form a point set in m-dimensional space, with each of the point in it describes one block of image x. As a result, we can calculate the distance between two images xi and xj as distance between the two point sets Si and Sj . We use the mean of distance between each of the K points in Si and Sj as the distance measure. This measure has been widely used in cluster methods, and proved to be robust to noise. The distance between images xi and xj in ICA space can be formulated as: d(xi , xj ) =

K K 1  i j (Sl , Sm ) K2 m=1

(9)

l=1

j Where Sli denotes the l-th column of Si , Sm denotes the m-th column of Sj , and (·, ·) denotes inner product of two vectors. After some easy formulation, we can simplify the distance to

1 sum(Si SjT ) (10) K2 where sum(·) denotes the sum of all the elements of a matrix. It is easy to see that this distance measurement is quite computationally efficient compared to L2 norm distance between Si and Sj . Under the assumption that the images lay on smooth manifolds embedded in image space, and the labeled data is limited, we use a semi-supervised algorithm L to learn the hypothesis in each view. The original method proposed in [23] is as follows: T Given a set of point X = {x1 , · · · , xq , xq+1 , · · · , xn }, f = [f1 , · · · , fn ] denotes a ranking function which assigns to each point xi a ranking value fi . The vector T y = [y1 , · · · , yn ] is defined in which yi = 1 if xi has a label and yi = 0 means xi is unlabeled. A connected graph with all the images as vertices is constructed the edges

and are weighted by the matrix B where Bij = exp −d2 (xi − xj ) 2σ 2 if i = j and Bii = 0 otherwise. d (xi − xj ) is the distance between xi and xj . B is normalized d(xi , xj ) =

456

F. Wang and Q. Dai

by S = D−1/2 BD−1/2 in which D is a diagonal matrix with its (i, i)-element equal to the sum of the i-th row of B. All points spread their ranking score to their neighbors via the weighted network. The spread is repeated until a global stable state is achieved. This label propagation process actually minimizes an energy function with a smoothness term and a fitting term. The smoothness term constrains the change of labels between nearby points, and the fitting term forces the classifier not to change too much from the initial label assignment. It has been proved that this iteration algorithm has a closed form of solution f ∗ = (I − αS)−1 y to directly compute the ranking scores of points [22]. From this formula we can discover that the initial value f0 has no effect on the final result, which is solely determined by y, S and α. Down to the case of our problem, there are another two issues to take into consideration. Firstly, the scale of our problem is very large, so we prefer using iteration algorithm, instead of direct inverse, which is more time consuming. Our experiment shows that a few iterations are enough to converge and yield high quality ranking results. Secondly, at the beginning of learning in one view, all the examples have been assigned ranking scores by the other view. The examples tending positive have values close to +1, while those tending negative have values near -1. In these scores, some could be changed, but those marked as +1 or -1 by the user in relevance feedback should not be changed since they are absolutely fixed. That means we have prior knowledge about the confidences of the labels proportional to their respective absolute values. Considering that yi stands for whether the example has a label in the standard semi-supervised algorithm, T which can also be regarded as the confidence, we set y = [y1 , · · · , yn ] as the ranking scores obtained from the other view. Since initial f0 is not crucial in iteration, it can also be set as equal to y at the beginning. Based on the predefined parameters, iterate f (t + 1) = αSf (t) + (1 − α) y for several times, Here alpha is a parameter in (0, 1), which specifies the relative contributions to the ranking scores from neighbors and the initial ranking scores. At last, each point xi is ranked according to its final ranking scores fi∗ (largest ranked first). The result of the propagation fi∗ is normalized separately as the h1 (x) or h2 (x) mentioned above in Section 3, which gives the probability that the sample is positive in separate views. Then we can deduce the disagreement of them by simply calculate their difference.

5

The Combined Scheme for the Proposed CBIR System

The integrated framework will be described in this section. First, the positive image set R+ is initialized as only the query image and the negative set R− as empty. The labels of all the images are initialized as zero. The times for relevance feedback is set as N . Other symbols are defined in Section 3. The following steps are performed:

A New Multi-view Learning Algorithm Based on ICA Feature

457

(1) On the positive set R+ , do ICA expansion and the basis vectors are obtained; (2) Based on general low-level feature, for each image xi ∈ R+ , we find its k-nearest neighbors Ci = {y1 , y2 , · · · , yk }, then we get the candidate image set C = C1 ∪C2 ∪· · ·∪C|R+ | ∪R+ ∪R− . T and U are labeled and unlabeled examples in C, respectively, that is, C = T ∪ U . The labels of images in R+ are changed to +1, and those in R− to -1; (3) Run Co-EM(L,V1,V2,T,U,k) in candidate set C for k times to learn h1 and h2 ; L is the algorithm proposed in Section 4 and Co-EM can be referred to Section 3. A typical value of 5 for k is enough to let the Co-EM algorithm converge to a stable point; (4) Sort the examples x ∈ U according to the absolute value of (h1 (x)−h2 (x)), those with large values are defined as contention points, that means, the two views are less confident of the labels of these examples. Select several examples with the largest value among contention points and ask user to label them; (5) The positive examples newly labeled by user are removed from U to R+ , and the negative ones to R− ; (6) N = N − 1. if N > 0, return to step (1); (7) Sort the examples according to h1 + h2 in descending order, and the final retrieval results are returned as the first several examples with largest value of h1 + h2 , that means, the two views both have high confidence on those examples. The candidate set is necessary when the whole image database is so large that the computation in the whole set will be time-consuming and needless. Additionally, in each iteration, some new examples are added into positive set, so there is no need to recalculate the basis vectors. When we do the ICA expansion, the de-mixing matrix W can be initialized as the matrix obtained in the previous iteration, and updated only by the subimages sampled from the newly added examples in positive set. This incremental learning advantage benefits from the characters of ICA, and guarantees the speed of our system.

6

Experiments and Discussions

The image database used in our experiments includes 5000 real-world images from Corel gallery. All the images belong to 50 semantic concept categories and 100 images in each category. The following features, which are totally 515 dimensions, are adopted as the general low-level feature: the 256-dimensional color histogram in HSV color space; the 9-dimensional color moments in LUV color space; color coherence in HSV space of 128-dimension; the 10-dimensional coarseness vector; 8-dimensional directionality; and the wavelet texture feature, 104 dimensions. To investigate the performance of our system , the following three algorithms are implemented and compared: (a) Our proposed multi-view learning algorithm, one view is ICA feature and the other is general low-level feature;

458

F. Wang and Q. Dai

(b) Our proposed multi-view learning algorithm, the two views are both general low-level feature; (c) Combine ICA feature and general low-level feature together as a single view, just adopt the semi-supervised algorithm proposed in Section 4. Each experiment is performed for 500 times, 10 times in each category. To simulate the real query process, the images are randomly selected from each category as the queries. The number of feedback rounds is set as 4 and in each round 3 images are returned as contention points for labeling. Here the system makes the judgement and gives the feedback automatically to simulate the user’s action. The retrieval accuracy is defined as the rate of relevant images retrieved in top 20 returns. Whether two images are relevant or not is determined automatically by the ground truth. The final averaged accuracy of retrieval results are shown in Fig.1, from which we can conclude that, our method (a) outperformed the other two experiments. The first point on each curve represents the accuracy obtained in the first round before any relevance feedback. As the round of feedback increases, the retrieval accuracy is getting higher. One point that has to be mentioned is that, the number of images for labeling and the round of feedback needed in our experiments are so small that it won’t make the user feel boring to make labels. Additionally, the time spent in retrieval is about 10s in a PC of P4 2.0GHz CPU and 1G RAM with M atlab implementation, which would be accepted by most users. To make a detailed discussion, we analyze the results in the following two aspects: ICA Feature vs. General Low-level Feature In experiments (a) and (b), both of the mechanisms of CBIR are multi-view learning algorithm, but the features adopted are different. In (b), another set of general features replaces ICA feature as the other view. Since the general features are mostly concerning the statistical characteristics of the images, their interaction on each other is not so significant as that between ICA and general feature. This means, the two views in the multi-view learning algorithm should be less correlated to achieve better performance. Our method handles this problem well, because ICA feature is from the view of human vision, while general features is on the view of computer. Multi-view vs. Single-view Experiment (a) and (c) are based on exactly the same features, and in (c), the distance between two images is measured as weighted sum of the distance of general feature and that of ICA feature, defined in Section 4. The better retrieval performance of (a) shows that, providing the same features, it is better to divide them into two parts and use the multi-view learning algorithm than to simply combine them together. The reason is that, the two views will interact and mutually provide the information that the other is lack of. Another remarkable phenomenon should be pointed out is that, when the round of feedback is more than 2, the retrieval accuracy of experiment (a) and

A New Multi-view Learning Algorithm Based on ICA Feature

459

1 Experiment (a) Experiment (b) Experiment (c) 0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0

1

2

3

4

Fig. 1. Retrieval Results Comparison

(c) would be close. The reason is probably that, the features adopted in (a) and (c) are almost the same, so the information that we can ultimately utilize is almost the same. The interaction in multi-view learning only has effects at the first several rounds, and with the increase of rounds, the information provided by the two views has been almost mixed fully and the labels they provide will get close, then they may perform similarly as the system (c) with combined features. Therefore, we can infer that, even the mechanisms of CBIR system are different, the final retrieval result after sufficient feedback rounds will only be related to the features we adopted and the feedback information provided by user. And this conclusion can be interpreted by the information theory as well. Then the advantage of our proposed system in practical applications is that, we can achieve high retrieval accuracy in the first several feedback rounds, i.e., 2 rounds may be enough, which can significantly improve the efficiency.

7

Conclusions

We have proposed a multi-view learning framework of CBIR, which is further consolidated with the feature extracted by ICA. At first, it is proved in theory that the ICA feature can provide more information than the original general low-level features for it accords with human vision. In the second place, the advantages of ICA feature and general low-level feature are integrated to improve each other in the scheme of the multi-view learning algorithm Co-EMT. This dramatically reduce the time of relevant feedback by the users. An the end, the semi-supervised learning algorithm in a single view is designed according to the specialties of the labels and the needs of Co-EMT. Owing to the forenamed

460

F. Wang and Q. Dai

characteristics of our proposal, our experimental results demonstrate the outstanding retrieval performance.

Acknowledgements This work is supported by the Distinguished Young Scholars of NSFC (No.60525111), and by the key project of NSFC (No.60432030).

References 1. A.J. Bell and T.J. Sejnowski. An Information-Maximization Approach to Blind Separation and Blind Deconvolution. Neural Computation, 1995, Vol: 7, pp. 1129–1159. 2. A.J. Bell and T.J. Sejnowski. The ’Independent Components’ of Natural Scenes are Edge Filters. Vision Research, 1997, Vol. 37, No. 23, pp. 3327–3338. 3. A. Blum and T.Mitchell. Combining Labeled and Unlabeled Data with CoTraining. Proc. of the Conference on Computational Learning Theory, 1998, pp. 92–100. 4. J.F. Cardoso and B. Laheld. Equivariant Adaptive Source Separation. IEEE Trans. on Signal Processing, 1996, Vol. 45, No. 2, pp. 434–444. 5. D. Hubel. Eye, Brain, Vision. Scientific American Library, 1995. 6. J. Hurri, A. Hyvarinen, J. Karhunen, and E. Oja. Image Feature Extraction Using Independent Component Analysis. Proc. IEEE Nordic Signal Processing Symp., Espoo, Findland, Sept. 1996, pp. 475–478. 7. A. Hyvarinen. A Family of Fixed-point Algorithm for Independent Component Analysis. Int. Conf. on Acoustic, Speech and Signal Processing, 1997, pp.3917-3920. 8. A. Hyvarinen and E. Oja. A Fast Fixed-point Algorithm for Independent Component Analysis. Neural Computation, 1997, Vol. 9, No. 7, pp.1483–1492. 9. N. Katsumata and Y. Matsuyama. Similar-Image Retrieval Systems Using ICA and PCA Bases. Proc. of the International Joint Conference on Neural Networks, Montreal, Canada, July 31-August 4, 2005, pp. 1229–1334. 10. T.W. Lee, M.S. Lewicki and T. Sejnowski. Unsupervised Classification with NonGaussian Mixture Models using ICA. Adavances in Nueral Information Processing Systems, 1999. 11. I. Muslea, S. Minton and C.A. Knoblock. Active + Semi-Supervised Learning = Robust Multi-View Learning. In Proc. of the International Conference on Machine Learning, 2002, pp. 435–442. 12. I. Muslea, S.Minton, and C.A. Knoblock. Selective Sampling with Redundant View. Proc. of National Conf. on Artificial Intelligence. 2000, pp. 621–626. 13. K. Nigam and R. Ghani. Analyzing the Effectiveness and Applicability of CoTraining. Proc. of Information and Knowledge Mnangement, 2000, pp. 86–93. 14. B.A. Olshausen and D.J. Field. Sparse Coding with an Overcomplete Basis Set: A Strategy Employed by V1? Vision Research, 1997, Vol. 37, No. 23, pp. 3311–3325. 15. A.T. Puga. A Computational Allegory for V1. Proc. of the 2nd International Symposium on Image and Signal Processing and Analysis, 2001, pp. 639–644. 16. A. Smeulders, M. Worring, A. Gupta and R. Jain. Content-Based Image Retrieval at the End of the Early Years. Proc. IEEE Trans. Pattern Analysis and Machine Intelligence, 2000, Vol. 22, No. 12, pp. 1349–1380.

A New Multi-view Learning Algorithm Based on ICA Feature

461

17. K. Tieu and P. Viola. Boosting Image Retrieval. Proc. IEEE Conf. on Computer Vision and Pattern Recognitino, 2000, pp. 228–235. 18. S. Tong and E. Chang. Support Vector Machine Active Learning for Image Retrieval. Proc. ACM Multimedia, Ottawa, Canada, 2001, pp. 107–118. 19. N. Vasconcelos and A. Lippman. Learning From User Feedback in Image Retrieval Systems. Advances in Neural Information Processing Systems, Denver, Colorado, 1999. 20. W.E. Vinje and J.L. Gallant. Sparse Coding and Decorrelation in Primary Visual Cortex During Natural Vision. Science, Feb. 18, 2000, Vol. 287, pp. 1273–1276. 21. R. Yan and A.G. Hauptmann. Co-retrieval: a Boosted Reranking Approach for Video Retrieval. IEE Proc. Vision Image Signal Processing, Dec. 2005, Vol. 152, No. 6, pp. 888–895. 22. D. Zhou, O. Bousquet, T. Lal, J. Weston, and B. Sch¨ olkopf. Learning with Local and Global Consistency. Advances in Neural Information Processing Systems, 2003, vol. 16. 23. D. Zhou, J. Weston, A.Gretton, O.Bousquet, and B. Sch¨ olkopf. Ranking on data manifolds. Advances in Neural Information Processing Systems 16, Cambridge, MA, 2004. MIT press. 24. X. S. Zhou and T.S. Huang. Comparing Discriminating Transformations and SVM for Learning during Multimedia Retrieval. Proc. ACM Multimedia, Ottawa, 2001, pp. 137–146.

A New Multi-view Learning Algorithm Based on ICA ... - Springer Link

the image retrieval, and through comparison, the conclusion is made that the. ICA basis ..... iteration algorithm has a closed form of solution f. ∗ .... Proc. of the Conference on Computational Learning Theory, 1998, pp. ... Video Retrieval.

324KB Sizes 1 Downloads 197 Views

Recommend Documents

A New Multi-view Learning Algorithm Based on ICA ...
1. A.J. Bell and T.J. Sejnowski. An Information-Maximization Approach to Blind. Separation and Blind Deconvolution. Neural Computation, 1995, Vol: 7, pp. 1129–1159. 2. A.J. Bell and T.J. Sejnowski. The 'Independent Components' of Natural Scenes are

Augmented reality registration algorithm based on ... - Springer Link
CHEN Jing1∗, WANG YongTian1,2, GUO JunWei1, LIU Wei1, LIN JingDun1, ... 2School of Computer Science and Technology, Beijing Institute of Technology, Beijing 100081, China; .... One is degrees and the other is meters or ... these years many successf

A DNA-Based Genetic Algorithm Implementation for ... - Springer Link
out evolutionary computation using DNA, but only a few implementations have been presented. ... present a solution for the maximal clique problem. In section 5 ...

A reordered first fit algorithm based novel storage ... - Springer Link
context, we call edges resulting from the 1st phase, thus drawn below the numbers, ..... Networks, SAMSUNG Electronics Co., 1999. [6] Crozier S, Guinand P. ... Turbo Decoding [C]// 9th International Conference on. Electronics, Circuits and ...

Dictionary Learning Based on Laplacian Score in ... - Springer Link
plied in signal processing, data mining and neuroscience. Constructing a proper dictionary for sparse coding is a common challenging problem. In this paper, we treat dictionary learning as an unsupervised learning pro- cess, and propose a Laplacian s

Asymptotic tracking by a reinforcement learning-based ... - Springer Link
Department of Mechanical and Aerospace Engineering, University of Florida, Gainesville, FL 32611, U.S.A.;. 2.Department of Physiology, University of Alberta, ...

Asymptotic tracking by a reinforcement learning-based ... - Springer Link
NASA Langley Research Center, Hampton, VA 23681, U.S.A.. Abstract: ... Keywords: Adaptive critic; Reinforcement learning; Neural network-based control.

Contrasting effects of bromocriptine on learning of a ... - Springer Link
Materials and methods Adult male Wistar rats were subjected to restraint stress for 21 days (6 h/day) followed by bromocriptine treatment, and learning was ...

A Linear Time Algorithm for the Minimum-weight ... - Springer Link
In this paper, we study the minimum-weight feedback vertex set problem in ...... ISAAC'95 Algorthms and Computations, Lecture Notes in Computer Science,.

A Linear Time Algorithm for the Minimum-weight ... - Springer Link
For each Bi, 1 ≤ i ≤ l, applying Algorithm II, we can compute a collection of candidate sets. FBi u. = {FBi .... W. H. Freeman and Company, New York, 1979.

LNAI 4285 - Query Similarity Computing Based on ... - Springer Link
similar units between S1 and S2, are called similar units, notated as s(ai,bj), abridged ..... 4. http://metadata.sims.berkeley.edu/index.html, accessed: 2003.Dec.1 ...

Interactive Cluster-Based Personalized Retrieval on ... - Springer Link
consists of a good test-bed domain where personalization techniques may prove ... inserted by the user or implicitly by monitoring a user's behavior. ..... As the underlying distributed memory platform we use a beowulf-class linux-cluster .... Hearst

Are survival processing memory advantages based on ... - Springer Link
Feb 8, 2011 - specificity of ancestral priorities in survival-processing advantages in memory. Keywords Memory . ... have shown that survival processing enhances memory performance as measured by recall, ..... Nairne, J. S., Pandeirada, J. N. S., & T

LNCS 4261 - Image Annotations Based on Semi ... - Springer Link
MOE-Microsoft Key Laboratory of Multimedia Computing and Communication ..... of possible research include the use of captions in the World Wide Web. ... the Seventeenth International Conference on Machine Learning, 2000, 1103~1110.

Interactive Cluster-Based Personalized Retrieval on ... - Springer Link
techniques based on user modeling to initiate the search on a large ... personalized services, a clustering based approach towards a user directed ..... Communities on the Internet Using Unsupervised Machine Learning Techniques. ... Distributed Compu

Are survival processing memory advantages based on ... - Springer Link
Published online: 8 February 2011. © Psychonomic Society, Inc. 2011 ... study investigated the specificity of this idea by comparing an ancestor-consistent ...

Fragments of HA based on Σ1-induction - Springer Link
iIΣ1 of IΣ1 (in the language of PRA), using Kleene's recursive realizability ... L(PRA), where in the presence of coding functions we may assume that no two .... q-realizes φ) by induction on the formation of φ. s always denotes one of r and q.

LNCS 4261 - Image Annotations Based on Semi ... - Springer Link
Keywords: image annotation, semi-supervised clustering, soft constraints, semantic distance. 1 Introduction ..... Toronto, Canada: ACM Press, 2003. 119~126P ...

Hooked on Hype - Springer Link
Thinking about the moral and legal responsibility of people for becoming addicted and for conduct associated with their addictions has been hindered by inadequate images of the subjective experience of addiction and by inadequate understanding of how

A Velocity-Based Approach for Simulating Human ... - Springer Link
ing avoidance behaviour between interacting virtual characters. We first exploit ..... In: Proc. of IEEE Conference on Robotics and Automation, pp. 1928–1935 ...