Automatic Name-Face Alignment to Enable Cross-Media News Retrieval

1 Introduction

With the explosive growth of multimodal news available both online and offline, how to integrate multimodal information sources to achieve more accurate cross-media news retrieval becomes an important research focus [Datta et al., 2008]. Usually multimodal news is exhibited with the form of captioned news images, which mostly describe stories about people [Liu et al., 2008]. Thus enabling automatic

porting cross-media news retrieval. However, because of the semantic gap [Fan et al., 2012], there may exist huge uncertainty on the correspondence relationships among names (in captions) and faces (in images) [Deschacht et al., 2007], [Yang et al., 2009], e.g., the relationship between names and faces in a news is many-to-many rather than one-to-one.

To achieve automatic name-face alignment, three inter-related issues should be addressed simultaneously: 1) multimodal analysis of cross-media news to identify better correspondences between names and faces; 2) discovery of the missing information for rare name; and 3) multimodal optimization to achieve more effective alignments for all possible name-face pairs. To address the first issue, it is very important to develop robust algorithms that are able to achieve more accurate name-face alignment. To address the second issue, it is very interesting to leverage large-scale Web news for missing information prediction. To address the third issue, it is critical to develop a new optimization algorithm with high accuracy rate but low computational cost.

Based on these observations above, a novel scheme is developed in this paper for facilitating automatic name-face alignment to enable more effective cross-media news retrieval. Our scheme significantly differs from other earlier work in: a) Salient names and cohesive faces are extracted from cross-media news. Such names might have higher correspondence possibility with the faces in news images, and such faces might have higher coherence with the salient names. b) Web news is integrated for discovering missing information for rare name. c) An efficient measurement and optimization mechanism is established to verify the feasibility of our automatic alignment algorithm. d) A new algorithm is developed to achieve more precise characterization of the correlations between names and faces. Such cross-media alignment is treated as a problem of bi-media semantic mapping, and modeled as a correlation distribution over semantic representations of names and faces. Our experiments on a large number of public data from Yahoo! News have obtained very positive results.

2 Related Work

Alignment of names and faces in cross-media news is not a novel task [Satoh et al., 1997&1999], [Berg et al., 2004&2005]. Earlier research considered acquiring the relevant faces based on the original text query over image captions, and ranking or filtering the returned images by a face detector. However, the irrelevant or incorrect results may be produced by utilizing the simple matching between a query name and captions. Most face recognition approaches are only applied to the controlled setting and limited data collections [Mensink et al., 2008], [Bozorgtabar et al., 2011]. Thus the increasing research interest focuses on combining textual and visual information to support precise alignment.

In recent years, there is some related research work by using textual information that accompanies the image. Yang et al. [2004] proposed an approach for finding the specific persons in broadcast news videos by exploring various clues such as names in the transcript, face, anchor scenes and the timing pattern between names and people. Everingham et al.

[2006], [2009] investigated the problem of automatically labeling appearances of the names in TV or film, and demonstrated that high precision can be achieved by combining multiple sources of information. Ozkan et al. [2006] proposed a graph-based method to retrieve the correct faces of a queried person using both text and visual appearances. Le et al. [2007], [2008] introduced an unsupervised method for face annotation by mining Web, which aimed to retrieve relevant faces of one person by learning the visual consistency among results retrieved from text-correlation-based search engines. It's necessary to mention the most interesting research work developed by Berg et al. [2004], [2005], [2007], Guillaumin et al. [2008], [2012], and Pham et al. [2010], [2011]. Berg et al. proposed a method for association of names to faces using more realistic dataset. Guillaumin et al. considered two scenarios of naming people in databases of news photos with captions, that is, finding faces of a single query person and assigning names to all faces. Pham et al. reported their experiments on aligning names and faces as found in the images and captions of online news websites.

Unfortunately, all these existing approaches have not provided good solutions for the following important issues:

(1) Analysis and Measure for Name Salience and Name-Face Cohesion in Deep Level - Most existing alignment methods focus on exploiting all the inter-linkage information between each name and each face in a same captioned news image, which leads to a serious combinatorial problem for possible name-face pairs. Without sufficient analysis of name salience and measure for name-face cohesion, some names and faces may form the noisy information and cause the insignificant name-face matching judgement.

(2) Discovering Extra Multimodal Information for Rare Name - Most existing work concerns finding faces of a specific name underlying such a precondition that the relevant face set consists of a large group of highly similar faces for the original name query. With the increasing growth of Web news, it appears that such Web information can mitigate the lack of relevant faces for rare name. According to our best knowledge, no existing research has made full use of Web news to achieve more accurate alignment for rare name.

(3) Multimodal Alignment Optimization for Name-Face Pairs - Finding the best matching between names and faces for all the pairs is very difficult and complex, and may become a NP-hard problem. Probability-statistical models can be used for computing the alignment likelihood, but have the serious problem of getting stuck in a local maximum. From the view of combinatorial optimization, it's a very significant way to set a global objective function, the local constraint conditions and an integer programming model.

3 Name Salience Identification

In a news caption, not all names are equally important. Thus it's necessary to evaluate every name in the same caption and judge which names may possibly co-occur with the faces in the original image-caption pair. Here, "salience" is defined as a measurement for the important degree of each name in a caption. A pattern for computing such a score is constructed based on the in-depth multi-level analysis of caption.

In the initial name list, the names are enumerated without any ranking. Since the syntactic structure is indicative of various information distribution in a caption, the relative rank of names can be confirmed by means of the roles of names and the relationships among names in the syntactic parse tree. (1) Syntactic Parse Tree Depth (SPTD) - This factor indicates the minimum depth for all the names of a certain name clustering in the parse tree for a caption. If a name has a shallower depth, it will have the higher importance. Given a caption with N names, each name is associated with its corresponding name clustering NCi and NCij represents the j th name in NCi. The depth value for each NCi can be defined as:

, . , . , . Ψ β ij NC SNC j i NC Depth SPT NC SPTD i _ min 1 _ _ (1) where SNC(NCi) denotes the size of NCi, i.e., the number of names with the co-reference relationship contained in NCi, which can also be understood as the occurrence frequency of the same names with different forms in the given caption; and SPT_Depth(NCij) is the depth of NCij in the parse tree.

(2) Syntactic Parse Tree Traversal Order (SPTTO) - This factor indicates the minimum breadth-first traversal order for all the names of a certain name clustering in the parse tree for a caption. It's proposed based on such a fact that among the nodes in the same level, a node with the higher priority of the traversal order means its higher relative importance. The traversal order value for each NCi can be defined as:

, . , . , . Ψ β ij NC SNC j i NC Order BFT SPT NC SPTTO i 2 _ _ _ min 1 (2) where SPT_BFT-Order(NCij) is the breadth-first traversal order of NCij in the syntactic parse tree.

In most cases the higher frequency for a specific name in a caption indicates its higher importance. It's very necessary to take such frequency into account as a key reference factor. Given N different names in a caption, in which each name corresponds to a name clustering NCi, i=1, …, N, the initial rank value, Relative Salience (RS), can be computed as:

where α, β and γ are the relative importance coefficients, α+β+γ=1. Maybe there is an extreme situation that only one name is involved in the caption, it's insignificant to evaluate the relative salience for such a unique name, RS(NC1)=1.

It's rather crude to assume that every name in the caption appears in the image. Thus it's very significant to make the further refinement to filter the names with the extremely lower possibility to be pictured as the detected faces. A visual survey for name-face pairs with the association linkage learned that the number of available faces is generally less than or equal to the number of detected names in a pair. We make the simplifying assumption that all the names with the initial rank values below a certain threshold may not be pictured as the selected faces. Given F faces and N names (F<=N) in the initial rank list for a pair, F' selected faces (F'<=F) are retained after face filtering and N names are refined to N' names (F'<=N'<=N) according to the visual exhibition information. Under such refinement, multiple names with the relatively higher importance are chosen to interpret the selected faces in the image, i.e., name denoising.

4 Name-Face Cohesion Measure

The general assumption for name-face alignment is that the appearances of the faces for a person/name are more similar than those for different persons/names. Thus it's very useful to establish a cohesion measurement to evaluate the intrinsic association degree among all the faces in the whole database.

Firstly, the similarity between two faces is evaluated based on the common neighbors among k-nearest neighbors for every face. Given two faces Fi and Fj in the local face set FS_Nm related to a real name Nm, if Fi is close to Fj and they are both close to FS_Nm, Fi and Fj are close with higher confidence for the mutual density since their similarity is determined based on all the faces in FS_Nm. Thus the similarity between a pair of faces can be defined as:

k where KNS(Fi, FS_Nm, k) and KNS(Fj, FS_Nm, k) denote the k-nearest neighbors when Fi or Fj positions in FS_Nm; and k is a dynamically varied value and changes according to the size of the local face set related to the current name.

Secondly, the Local Density Score (LDS) is utilized to measure the density for each face in the local face set. The larger the density is, the more relevant the face and the name associated with the face set are. The LDS for a face Fi can be described as the average similarity between this face and its k-nearest faces in the same face set. The higher LDS indicates the higher connectivity between Fi and its k-nearest faces, that is, Fi is more relevant with Nm associated with FS_Nm.

Thirdly, the Local Cohesion Degree (LCD) is especially proposed to measure the mutual density among all the faces in the face set related to a particular name under the current global alignment manner, and defined as:

If a face has the higher LDS value in the related face set, this face will have the higher density with the other faces in the same face set and then will be more relevant with the name corresponding to the face set. Therefore, as the sum of the LDS values for all the faces in a local face set, LCD can be well representative of the mutual density among all the faces in this face set. The larger the LCD is, the higher the mutual density among all the faces in the face set will be. The higher LCD can be regarded as a good indication of the higher local cohesion for the face set, that is, the face set is more relevant with the assigned name. Under an arbitrary global alignment manner, the sum of the LCD values for all the face sets can reflect the whole global cohesion for the complete face set related to all the names. Thus the larger sum of the LCD values for the face sets embodies the higher global cohesion for the complete face set, and then further expresses the higher global relevance between the whole face set and the name set under the current global alignment manner.

5 Web Mining for Rare Name

The name-face alignment can also be regarded as a restricted face naming or retrieval problem. The general assumption is that the number of faces corresponding to the query name is relatively large. However, some rare names may lead to such a fact that the number of their faces is very small. The "rare name" here does not mean the real single name, but refers to the rare person that occurs only several times or even one time in the whole database. Thus for the alignment evaluation of rare name, it's necessary to establish a discovery mechanism to supplement more available multimodal information.

As the vast knowledge base, Web has become the best resource to create a reference annotated face image set for rare name. Face images about a specific person can be automatically retrieved by using Web image search engines. Based on Web mining, a certain number of relevant face images and their annotations can be discovered simultaneously, and become the valuable enrichment for rare name. All the name expression forms for a rare name are collected in a rare name clustering. Each form in the clustering is taken as an original name query and submitted to Web image search engines. Thus a list of relevant face images and their annotations can both be acquired. Each returned face image will be processed through the face detection and filtering, and the available and meaningful face images are kept.

Given a rare name Name0, all of its face images can be ranked according to the salience value of Name0 in each face image. The Top-R face images are maintained as the complementary face information to Name0. R is a generalized threshold for the complementary face image selection, and does not have a fixed setting. Because there may be different numbers of face images returned for different rare names, R means a limited proportion of the face images preferred, and varies with the size of the returned face image set.

6 Name-Face Alignment via ISSAGA

We consider such an alignment task as an optimization problem for searching the optimal name-face matching in each pair. The Standard Genetic Algorithm (SGA) can obtain the global optimal search using genetic operations, but there are some disadvantages such as "prematurity" and the poor search power for local optimal solutions [Wang et al., 2005]. However, the Simulated Annealing Algorithm (SAA) has the stronger ability for local optimization [Andresen et al., 2008]. Thus the Improved Self-Adaptive Simulated Annealing Genetic Algorithm (ISSAGA), which integrates the advantages of SGA and SAA and a self-adaptive probability adjustment strategy, is introduced to make a better solution.

6.1 Multimodal Name-Face Alignment

Suppose P image-caption pairs include F real faces (i.e., Not Null Face) and N real names (i.e., Not Null Name), PS denotes the Image-Caption Pair Set for P pairs; FS denotes the Face Set for F faces; NS denotes the Name Set for N names; W_FPij denotes whether the face Fj belongs to the pair Pi, Pi PS, Fj FS, i=1, …, |PS|, j=1, …, |FS|, which is equal to:

W_NPik denotes whether the name Nk belongs to Pi, Pi PS, Nk NS, i=1, …, |PS|, k=1, …, |NS|, which is equal to:

FPi denotes the face set for Pi, FPi={Fj|W_FPij=1, Fj FS}, Pi PS, i=1, …, |PS|; NPi denotes the name set for Pi, NPi={Nk|W_NPik=1, Nk NS}, Pi PS, i=1, …, |PS|; W_FNkj denotes whether Fj is assigned to Nk, Fj FS, Nk NS, j=1, …, |FS|, k=1, …, |NS|, which is equal to:

FS_Nk denotes the total associated face set for Nk based on the global assignment of names and faces, Nk NS, k=1, …, |NS|; LCD(FS_Nm, k) denotes the mutual density among all the faces in the face set FS_Nm related to the name Nm under a certain global alignment manner, Nm NS, m=1, …, |NS|. Thus our mathematical model can be formalized as:

face set included in each pair is the universal face set of the whole database, and the union set of the name set included in each pair is the universal name set. 2) FPi∩FPj=Ф, i≠j, i, j=1, …, |PS|. That is, there is no intersection set between the face sets of two pairs. 3) |FPi|=|NPi|, i=1, …, |PS|. That is, to establish more efficient coding information for the optimization algorithm, we form a constraint that the numbers of faces and names in an arbitrary pair must be same. Null faces or names can be used to supplement the less ones. 4) 1 _ 1 η ⎠ _ i j NP k k FN W , Fj FPi, Nk NPi, i=1, …, |PS|, j=1, …, |FPi|. That is, in a pair each real face can only be assigned to at most a real name in the same pair. 5) 1 _ 1 η ⎠ _ i j FP j k FN W , Fj FPi, Nk NPi, i=1, …, |PS|, k=1, …, |NPi|. That is, in a pair each real name can only be assigned to at most a real face in the same pair. 6) 0 _ , _ ⎠ ∠ ∨ i l i j j NP N FP F l FN W , i=1, …, |PS|, j=1, …, |FPi|, l=1, …, |NS|. That is, the face that belongs to a pair can only be assigned to the name in the same pair.

6.2 ISSAGA-based Multimodal Optimization

a. Encoding of Chromosome

The names from all the pairs are positioned in a fixed order, and the faces from all the pairs are ranked in segments and then grouped into a chromosome. Each chromosome corresponds to a solution. A repeatable natural number encoding pattern is adopted to design the gene gij of the chromosome C, C={gij}, i=1, …, |PS|, j=1, …, |FPi|, where i is the pair number; j is the face number for the pair Pi; and gij denotes the number for the face Fj in Pi. C can be expanded as {g11, …, g1|FP1|, …, gi1, …, gi|FPi|, …, g|PS|1, …, g|PS||FP|PS||}, where {gi1, …, gi|FPi|} is a segment and each segment keeps the relatively independent relationship with the other segments.

b. Initial Population

The initial population P(t) with L chromosomes is produced by a stochastic method. The stochastic order in each segment is used to generate an initial chromosome C, and then the other L-1 different chromosomes are achieved from C. According to the objective function value, the current best chromosome can be regarded as the initial optimal solution.

c. Self-Adaptive Selection-Reproduction Operator

According to our mathematical model, the objective function for the l th chromosome in a population can be constructed as:

To limit the quantity of chromosomes that dissatisfy the constraints in the new population, we use a Multiple Roulette Wheel method and compute the selection probability as:

where M is the number of chromosomes in the current population, and f'( ) is obtained through the self-adaptive transformation based on the original fitness function f( ) to avoid "prematurity" and "slow convergence speed".

where fmax and fmin are the maximum and minimum fitness values for the current population; g is the number of generations under the current genetic manner; gmax is the maximum number of genetic generations; and a is a constant parameter.

d. Self-Adaptive Simulated Annealing Crossover Operator Considering in most cases there are two or three faces in a news image and the maximum order number for the faces in a segment of a chromosome is also two or three, thus a One-Point crossover method is adopted. Two special strategies of the self-adaptive crossover probability (PC) computation and the simulated annealing operation are introduced.

where favg is the average fitness value in the current population; PC1 and PC2 are two predefined parameters. The simulated annealing operation is utilized to judge whether the inferior solution should be accepted to substitute the previous chromosome when facing with an inferior solution and generate a probability (PA) for accepting the inferior solution.

where f(Cl') denotes the fitness value for the new chromosome Cl' generated after the crossover; T0 is the initial temperature for the simulated annealing operation; and δ is a preset ratio coefficient for the temperature reduction.

e. Self-Adaptive Simulated Annealing Mutation Operator The mutation operator cannot be executed among segments but inside every segment by utilizing an Exchange Mutation method. Meanwhile, two special strategies of the self-adaptive mutation probability (PM) computation and the simulated annealing operation are also introduced.

where PM1 and PM2 are two predefined parameters. Given a chromosome Ck={S1, …, Si, …, SN}, N=|PS|, Si is the i th segment. Taking a segment as a unit, the 2-exchange mutation is accomplished among the segments of Ck by PM(Ck).

7 Experiment and Analysis

7.1 Dataset and Evaluation Metrics

Our dataset is established based on Labeled Yahoo! News Data constructed by Berg et al. [2005] and further developed by Guillaumin et al. [2008]. It consists of 20,071 captioned news images, and all the image-caption pairs contain 31,147 face images that belong to 5,873 different persons/names. To evaluate the effectiveness of our algorithm, the ground truth associations are considered to measure the benchmark metrics for name-face aligned pairs. Correct Label Rate (CLR) is defined as the percentage of correct pairs generated in the ground truth links. Our evaluation can not only consider the association links between real names and faces, but also concern the assignments to null names or null faces.

7.2 Experiment on Cross-Media Alignment

Our cross-media alignment model is created by integrating Name Salience Ranking (NSR), Name-Face Cohesion Measure (NFCM), Web-based Multimodal Information Mining (WMIM) for rare name and the Improved Self-Adaptive Simulated Annealing Genetic Algorithm of ISSAGA. To investigate the effect of each part on the whole alignment performance, we introduce three evaluation patterns: 1) Baseline(NFCM) - based on the basic preprocessing with No NSR for captions and NFCM for images; 2) Baseline(NFCM)+NSR - based on the baseline model with NFCM, NSR is integrated; 3) Baseline(NFCM)+NSR+WMIM - by fusing the baseline model with NFCM and NSR, WMIM is added. These patterns are tested in two alignment optimization manners of general GA and ISSAGA, which aims to explore the advantages of our proposed ISSAGA.

To verify the consistency of our alignment model in different mapping cases, the performance evaluation focuses on two assignment schemes. One contains the specific mappings of Real Face->Null Name (RF->NN) and Real Name->Null Face (RN->NF) except the general mapping of Real Face<->Real Name (RF<->RN), and another only involves RF<->RN. The experimental results are shown in Table 1.

It can be seen from Table 1 that for the name-face linking in our whole dataset, we can still obtain the best CLR value of 65.77% in the same evaluation pattern of fusing Baseline with NFCM, NSR, WMIM and ISSAGA and excluding the mappings related to Null Name and Null Face. In comparison with the baseline model, the alignment performance could be promoted to a great degree by successively adding NSR and WMIM under our ISSAGA-based optimization. Compared the results for different assignment schemes, we can observe that when only considering the alignment between real names and faces, the alignment performance increases and becomes pretty good. Although when evaluating the links with null names and faces the alignment performance maybe slightly influenced by such extra uncertainty, the CLR can still reach a relatively high value. Through the comparison between the results based on the traditional GA and ISSAGA, it can be found that our ISSAGA is obviously superior to the traditional GA and more suitable for the alignment optimization.

In addition, the convergence curves of the best solutions found by the traditional GA and our ISSAGA are shown in Figure 1. The traditional GA improves the solutions very fast within 60 iterations, and reaches a plateau after that. In contrast, our ISSAGA keeps constantly improving the solutions with more iterations, gives much better solutions, and has a stronger ability of escaping from the local optima. It can be concluded that our ISSAGA provides better solutions within a satisfactory amount of time and is very useful to contribute better solutions for such a special optimization problem.

7.3 Experiment on Cross-Media News Retrieval

To further explore the applicability of our alignment model in cross-media news retrieval, we particularly select 16 names with the larger number of faces in the whole dataset as the query topics and carry out the retrieval runs based on the original dataset and the dataset with the alignment processing respectively. The Precision (P), Recall (R) and F-measure (F) values for each name query are shown in Table 2. It can be seen that the best run is based on the dataset with alignment processing, and its results exceed those by another run based on the original dataset without any alignment processing dramatically. For all the name queries, their F values under the run with alignment processing are obviously higher than those under another run, and the highest increase can reach to 20%. By adopting our alignment model for the original dataset, the whole retrieval performance for the selected name queries has gained the significant improvement.

7.4 Comparison with Existing Approaches

To give full exhibition to the superiority of our alignment model, we have also performed a comparison between our approach and the other classical ones in recent years. Three approaches developed by Berg et al. [2005], Guillaumin et al. [2008] and Pham et al. [2010] are analogous with ours, and then we accomplished them on the same dataset. The experimental results are presented in Table 3, which reflect the difference of power between these four approaches.

It can be found from Table 3 that for the name-face linking on our whole dataset by Berg/Guiilaumin/Pham et al.'s approaches, we can obtain the best CLR values of 52.71%, 54.41% and 55.61% in the evaluation patterns of Berg/Guill/ Pham [No NSR] and excluding the mappings related to Null Name and Null Face respectively. The main reason is that when only considering the alignments between real faces and real names, all these three approaches can reduce the combination ambiguities between names and faces in the same image-caption pair to a certain degree and then the relatively better CLR values can be obtained. Compared the results of Berg/Guill/Pham [No NSR] and our baseline model with NFCM, we can find the best CLR value of 56.78% appears in the results of our model, which is obviously higher than those CLR values of 52.71%, 54.41% and 55.61% for Berg/Guill/ Pham [No NSR] respectively. This implies that both of our name-face cohesion measure mechanism NFCM and combination optimization algorithm ISSAGA are feasible for facilitating more effective name-face alignment evaluation and optimization. Furthermore, through fusing our NSR and WMIM with Berg/Guillaumin/Pham et al.'s approaches in turn, the better performance can be acquired, in which the best CLR values are considerably improved to 59.93%, 62.43% and 62.73% respectively. This confirms the prominent roles of our NSR and WMIM in automatic name-face alignment once again. Compared with the improved Berg/ Guillaumin/Pham et al.'s approaches that integrate with NSR and WMIM, our alignment model with NFCM, NSR, WMIM and ISSAGA can still present the better performance and the best CLR value of 65.77% differs greatly from those of Berg/Guillaumin/Pham et al.'s. This indicates that our approach is really superior to Berg/Guillaumin/Pham et al.'s, and also further confirms that our alignment model with NFCM, NSR, WMIM and ISSAGA is exactly a better way for determining name-face associations and can support such a special combinatorial optimization problem more effectively.

7.5 Analysis and Discussion

Through the analysis for the aligned name-face linkages with failure, it can be found: 1) The alignment acquisition is associated with the preprocessing for image-caption pairs, especially for caption text. It's easier for image and caption preprocessing to produce some errors or missing detections for faces and names. Such wrong or undetected information will seriously affect the whole alignment performance. 2) The other names that co-occur with a certain name in the same caption maybe helpful for facilitating this name's related face finding and naming. The same case is for faces in images. Such name or face co-occurrence can be utilized to further improve the alignment measurement. 3) For multimodal information from Web mining, instead of directly using the feedback from image search engines, it's helpful to exploit a relevance evaluation manner to select more appropriate complementary information for rare name. (4) It is a novel way to solve name-face alignment from the view of combinatorial optimization. Although our ISSAGA have obtained the satisfactory performance on large-scale dataset, making more highly specialized consideration about genetic operations will be more beneficial to acquiring more feasible solutions. 5) Some news captions involve extremely rare name. With very limited information supplement from Web, such news lacks enough multimodal information for alignment processing. This may be a more stubborn problem.

8 Conclusions and Future Work

A new algorithm is introduced in this paper to support more precise automatic name-face alignment for cross-media news retrieval. The multi-level analysis of image-caption pairs is established for characterizing the meaningful names with higher salience and the cohesion between names and faces. To remedy the issue of rare name, the extra multimodal information is acquired through Web mining. ISSAGA is introduced to solve the alignment optimization better. Thus a novel model is developed by integrating all these aspects to effectively exploit name-face correlations. Our future work will focus on making our system available online, so that more Internet users can benefit from our research.

Acknowledgments

This work is supported by National Natural Science Fund of China (No. 61170095; No. 61272285), National Science & Technology Pillar Program of China (No. 2012BAH59F04), 973 Program of China (No. 2010CB327906), Shanghai Outstanding Academic Leader Program (No. 12XD1400900), and Shanghai Leading Academic Discipline Project (No. B114). Cheng Jin is the corresponding author.