Writer Identification and Verification: A Review Siew Keng Chan1, Yong Haur Tay1 and Christian Viard-Gaudin2 1

Computer Vision & Intelligent Systems (CVIS) Group, Faculty of Information & Communication Technology Universiti Tunku Abdul Rahman (UTAR), MALAYSIA. [email protected] 2

IVC, IRCCyN, UMR CNRS 6597 Ecole Polytechnique de l'Université de Nantes, FRANCE. [email protected]

Abstract Writer identification and verification has been studied over the past decade due to the high demand of such implementations in a wide variety of offline handwritten data applications. Among the applications are verification of the identity of a person whom he or she claimed to be and forensic and historic document analysis in identifying the identity of the writer of the documents. This paper provides a review on the state-of-art of writer identification and verification techniques which have been used and implemented. There are three main stages involved in writer identification and verification: feature extraction phase, classification phase and verification phase. In each phase, commonly used techniques and algorithms would be discussed.

them by taking into account the temporal nature of data presents in online handwriting. For writer identification, there are two problems to be considered. They are text-independent and textdependent approaches [1]. Both approaches are different in terms of the handwritten document whether they contain the same text or not. Textdependent approach means that the system can only identify the writer for text which is composed of the same specific text which is shown in Figure 1 or there should be a recognizer which has been trained to identify certain writer. Signature verification is a special example of such approach whereby the text is the signature of a signer. For text-independent approach, identification can be carried out irrespective to the contents of the handwritten text. Thus, identification task can be carried out easily on any given documents.

Keywords: Writer Identification, Writer Verification, Feature Extraction, Classification Techniques

1. Introduction Writer identification deals with identifying the writer of a handwritten document. Each individual’s handwriting is unique and thus it is feasible for person identification purposes. This is rather a new area of research as compared to signature verification which has been researched in depth over the years [2]. Signature verification involves the process of verifying a signature whether it is genuine or forgery by comparing the signature with few templates stored in the database. Most of the present studies concerning writer identification deal with offline handwriting [3] [12] [16]. However, with the widespread of digital pen and paper solutions, it would be interesting to extend

Figure 1: These two handwritten lines are written by two different writers. Two different tasks can be identified: one is writer identification, another one is writer verification. In the identification process, it strives to identify the writer of the handwritten text by comparing the handwritten text with those stored in the reference database. The output would be a list of ranked handwritten texts starting with those having the highest similarity with the handwritten text. In the verification process, it aims at verifying the writer of the handwritten text. Generally, the handwritten text would be compared with another text to verify

whether both texts are written by the same writer. The text used in the comparison would be the highest ranked handwritten document extracted from the reference database in the identification process. From these two processes, it seems that the identification process acts as a filtering step for verification process [5] [6]. According to [1], handwritten text can be categorized into sub-character level, character level, word level, line level and paragraph level where each conveys different individuality information for identification purposes. Sub-character is also known as grapheme; it gives the information of the basic structure of a character. One specific character exhibits different allographs. One writer will use actually the same allograph most of the time, while another writer would probably use another allograph of this same character. Word tells the information about how the characters are connected. Line tells the arrangement of the entire sentence and paragraph gives the global information of few lines of text. From all these levels, one can study the features to be used for writer identification. For example, at the sub-character level, the commonly used feature is the grapheme codebook. Handwritten Documents

Preprocessing Phase

Feature Extraction Phase

Classification Phase

Verification Phase

Figure 2. The processes involved in writer identification and verification. In [1], it listed down few of the writer identification systems that have been carried out solving different kind of problems and implementing different kind of methods. It also listed the pros and cons of each of the writer identification systems. Traditionally, the identification of the writer consists of three processes [5] [12]. This is illustrated in Figure 2. The three processes are pre-processing phase, feature extraction phase and classification phase. The techniques used in feature extraction phase and classification phase would be discussed in the next section. Pre-processing phase will not be discussed in this communication.

2. Approaches and Methods This section summarizes few techniques and methods for feature extraction phase, classification

phase and verification phase that have been implemented in writer identification.

2.1 Feature Extraction Phase Feature extraction phase is a process of identifying features that can be used in discriminating the writer in writer identification and verification. There are three groups of features in writer identification namely global measures, local measures and measures related to individual character shapes [10]. Eight features would be discussed here. They are run-length distribution, edge-based distribution, edge-hinge distribution, graphemes codebook, stroke-endings, autocorrelation and dynamic features. 2.1.1 Run-length Distribution. According to [11], a run is a connected sequence of pixels with the same value. The length of the run is computed by measuring the same pixel value either in vertical and horizontal direction. The accumulated run lengths for both horizontal and vertical run would then be normalized and put into a probability distribution function. This feature represents the probability of certain vertical or horizontal run lengths occur in the handwritten text of a writer. In [4], the considered run length size is up to 100 pixels with 120 pixels as the height of the written line. 2.1.2 Edge-based Distribution/Slant. This feature describes the orientation of strokes with respect to the baseline of the handwritten text as well as the regularity of handwriting through the computation of slant [11]. Each individual’s handwriting slant is different. Thus, it can be used as a feature to discriminate handwriting of different writers. In [4], the algorithm used to extract this feature from handwritten text is through the use of a convolution filter with two orthogonal Sobel kernels. This step is to produce a thinned version of the handwritten text. Then, a sliding window is used to compute the directions of the thinned text where only the ‘on’ pixels are considered. The central point in the sliding window is used to check for direction of the entire fragment of edge. The computed directions would then be tabulated into a probability distribution. If the probability distribution of slant shows higher peaks, it means there is a plot of regular handwriting and vice versa [11]. 2.1.3 Edge-hinge Distribution. Edge-hinge distribution is a new feature introduced by [4]. The difference of this feature with slant is that it considers two edges’ direction rather than one. It is a feature that considers the change of direction during writing [11]. The process of extraction is the same as slant. However, in the sliding windows, two edge

fragments’ directions exiting from the central point are considered which then give a joint probability distribution. This feature has been compared with other features (edge-based distribution, run-length distribution, graphemes codebook and direction cooccurrence distribution) and achieved the better result with 91% success identification and 4.8% Equal Error Rate (ERR) for verification procedure as compared to those mentioned [10].

Figure 3. This figure shows the graphemes codebook generated using K-means clustering and containing 400 graphemes. Figure was extracted from [10]. 2.1.4 Grapheme Codebook. Grapheme is a small segmented handwriting [11]. For grapheme extraction, two stages involved which are the grapheme codebook construction stage and grapheme feature construction stage [3] [10] [11]. In [10], this feature has been compared with other features as mentioned in edge-hinge distribution section and has achieved a result of 92% success identification and 5.8% of EER for the verification task. In the grapheme codebook construction stage, graphemes are extracted from samples of handwritten text. This step is performed by segmenting the handwriting text into segments of lines and then the segmented lines are segmented using vertical segmentation algorithm to obtain small handwriting segments. Each segment would contain zero, one or more than one grapheme. Graphemes are then extracted from these segments and processed using Moore’s algorithm [10] to obtain the contours of the graphemes. The collected graphemes are used to construct the codebook either using Kohonen SelfOrganizing Feature Map (SOFM) [3] [11], multiple fast sequential clustering algorithm [5] or k-means clustering [10]. The codebook is finally composed of classes of graphemes clustered accordingly and an example is shown in Figure 3. The codebook generated in the previous stage is used as a reference to build the probability density function for the writer. This stage is known as grapheme features construction stage [11]. The test handwritten text would be segmented accordingly as mentioned previously. The segmented graphemes

would then be compared to those stored in the codebook by computing the distance using Euclidean distance or other distance measures. The succeed matches would then be tabulated into probability density function which can be used for classification. 2.1.5 Stroke Endings. A stroke ending is defined as the end-point of a stroke which happens when a pen either has entered or left the paper during writing [11]. Sliding window is used to check for stroke ending with some constraints applied. When a stroke ending is found, the direction of the stroke ending is identified which then be tabulated into a probability density function. 2.1.6 Autocorrelation. Autocorrelation is a feature that tells the regularity of handwriting. The feature is computed by obtaining the dot product of the shifted lines of the original handwritten text with the original lines. The original lines of handwritten text is firstly shifted according to a fixed number of pixels and then it is multiplied with original handwritten text [4] [11]. The products are summed up and normalized to a value between 0 and 1. This value shows the regularity of handwriting [11]. 2.1.7 Dynamic Features. These dynamic features can only be found in case of online writer identification. Among the dynamic features are pressure, altitude and azimuth. These features have been used combined with static features (word shape) of handwriting by [2] to check the interdependency of both features. The reported False Acceptance Rate (FAR) was 8.6%. 2.1.8 Others. There are many other features which have been used in writer identification such as entropy [4], baseline direction [11], and number of black pixels [13]. Entropy relates to the ink distribution on the handwritten document [4] or a measure of degree of disorder [13]. Baseline direction is the regressed imaginary of a line of handwritten text obtained by using the minimum points. Number of black pixels is the count of the total number of black pixels in the handwritten document which gives the information of pressure, thickness of strokes and size of writing [13]. 2.1.9 Problems in Feature Extraction. Among the features listed, some have high discriminative power while the rest have lower discriminative power. For example, baseline direction is dependent on the paper of the handwritten text as most of the time the writer follows the line given and it can be easily forged by other writers [11]. Thus, this feature will not be suitable for writer identification. In case of dynamic features, it is only available in online handwritten texts. Thus, the system must be customized to either

online writer identification or offline identification systems. In the case of graphemes codebook, a lot of training data is needed to have a variety of graphemes which are different for each writer. This will eventually lead to expensive computation. Thus, features have to be studied to identify their discriminative power to achieve accuracy in writer identification and verification.

2.2 Classification Phase This part discusses the techniques used for writer identification. There are two major groups of classification techniques which are classifier, known as supervised classification technique and clustering, known as unsupervised classification technique. 2.2.1 Classifiers. Classifiers are commonly used in the writer identification phase where the handwritten text is first classified into few classes based on their similarity. The classifiers that would be discussed are K Nearest-neighbours Classifier, Bayesian Classifier, Weighed Euclidean Distance and neural networks. 2.2.1.1 K nearest-neighbours Classifier (K-NN). KNN is a simple density estimation technique [14]. This method was used in [10] with leave-one-out strategy. All the samples in the database are compared with the test handwritten document. The result would be a list of ranked document with increasing distances. In [12], this is used for computational simplicity. The feature vector of the test handwriting sample is compared with the ideal feature vector from each class in the training set using Euclidean distance. The minimum distance obtained from the computations would decide the test sample’s class. The classifier was tested on 40 different writer using features extracted from multichannel Gabor filtering technique and grey-scale cooccurrence matrix and the success identification result was 85%. 2.2.1.2 Bayesian Classifier. This classifier has been used by [9]. The classification strategy was leave-onout strategy. Weighted distance from the center of the cluster is the important criteria here. The center point of each cluster and the covariance matrix are computed for the training procedure leaving one point. However, due to expensive computational steps, the means and covariance matrix have been weighted. The achieved experimental result on a set of 50 writers for identification procedure was 92.48% for English words and 92.63% for Greek words. The tested data consists of a single word only. For verification procedure, the error rate was in the order of 5%.

2.2.1.3 Weighted Euclidean Distance Classifier (WED). This type of classifier has been used by [12]. There is a representative feature vector for each writer. The test sample is compared with a set of known writers’ feature vector and minimum distance would decide the class of the test sample. The classifier was tested on features extracted using multi-channel Gabor filtering technique and the greyscale co-occurrence matrix technique on 40 different writers and achieved a success identification result of 96.0%. 2.2.1.4 Supervised Neural Networks. In [13], multilayer perceptrons (MLP) with 11 inputs, 5 hidden and one output neurons, respectively, has been trained to classify the distances computed from features extracted from handwritten text (document level, paragraph level, line level and character level) into within-author distances or between-author distances. The system was tested using within-author distances and between-author distances on all the levels of features extracted and achieved average accuracy with 95%, 88%, 82.9% and 83.5% with respect to the feature level. In [9], the MLP used is 20-20-6 neural network. The input is a 20dimensional feature vector and the output is a six-bit binary number which gives the cluster which the test sample belongs to. The achieved experimental result is in the order 3.5% of identification error rate and in the order 2% of verification error rate. Hidden Markov Model (HMM) based recognizer was used in [8]. Each individual’s handwritten text lines would be used to train the recognizer which gives result of a recognizer for each individual. The output for the identification process is a list of ranked possible writers in decreasing order. The system was tested on a set of 100 writers with 96.56% of correct identification and over 8600 text lines from 120 writers with 2.5% of Equal Error Rate (EER) for verification procedure. 2.2.2 Clustering. Clustering techniques that would be discussed are K-means Clustering and Kohonen Self-Organizing Feature Map which are commonly used in clustering the training samples such as construction of graphemes codebook. 2.2.2.1 K-means Clustering. This technique has been used to accelerate convergence and simplify computation steps [14]. A point of representative is used to represent each cluster. As mentioned previously, this method has been used in clustering the graphemes or primitives from the handwritten text. In [1], this technique has been used to cluster the primitives (shape based curve) extracted from handwritten text into clusters. In [10], this clustering algorithm was applied to the training set of graphemes extracted from the handwritten samples to

construct the graphemes codebook. Those similar graphemes would be grouped in the same cluster with one representative in each cluster. 2.2.2.2 Kohonen Self-Organizing Feature Map (KSOFM). This has been used in clustering the graphemes extracted from handwritten texts. This clustering method was used to construct the grapheme codebook [3] [11]. Each node in a KSOFM is fed with input vectors and then they would produce different output as it depends on response of the node with the weight that is assigned in each of the nodes. The node with the highest response will update its weight accordingly at each loop so as its neighbour nodes. The size of the neighbourhood would decreases as iterations increased. The end result would be a map that maps the training data well. 2.2.3 Problems in Classification Phase. The major problem with classification phase is the expensive computation steps. Most of the methods require a great deal of training data to learn about the writers’ handwritten text. Since each individual writes differently, with the use of methods such as HMM, the need to have trained a recognizer for each individual would definitely need a sufficient amount of training handwritten samples. These samples must also cover all the possible handwritten texts that could be written by the writer. Clustering methods which are used in generating grapheme codebook would need even large sample of data from each writer. This is to be able to identify different sets of graphemes.

2.3 Verification Phase In this phase, it involves verifying the identity of the writer whether it is true against what he or she claims. The most commonly used method is through the calculations of a similarity score and this similarity score is compared with a threshold to make the decision whether the handwritten text is written by the claimed writer. Threshold is a value which is abstract and commonly determined empirically. This phase would generally lead to the evaluation of the system on a testing database to results of error rates. The commonly used error rates are False Rejection Rate (FRR), False Acceptance Rate (FAR) and Equal Error Rate (EER). FRR is the number of wrongly rejected genuine handwritten text. FAR is the wrongly accepted forgery handwritten text. EER is the equivalent point of FAR and FRR.

3. Discussion In this section, few issues related to writer identification would be discussed.

3.1 Dependent versus Independent As mentioned earlier, there are two problems to be considered which are text-dependent approach and text-independent approach. Text-independent approach would simulate the actual situations in the real world where different types of scripts and texts exist and suitable in identifying different documents’ writers. For text-dependent approach, it is the same as signature verification where same text or scripts are provided. Recognizers have to be trained using samples from the writer for identification and verification tasks.

3.2 Discrimination Power of Features Researches have been carried out to study the discriminative power of the features that are extracted from handwritten text [4] [13] [15]. For example, evaluation done in [4] [15] is to evaluate the performance of the edge-based features or angular features in comparison with non-angular features in writer identification. The result shows that angular features such as edge-direction distribution and edge-hinge distribution outperform non-angular features such as run-length distribution and autocorrelation [4] [15]. New feature has been proposed for writer identification i.e. edge-hinge distribution [4] which was tested and proven to improve the performance of writer identification [4]. However, there is no feature that can really claimed to have the highest discriminative power as it is subjected to the database of handwritten text that the features are extracted from and the database that is used in comparing the features to check the discrimination power. Combinations of features have also been tested in [10] to improve the performance of writer identification system. Working at the character level appears appealing but it requires an efficient segmentation and recognition system, which is quite difficult to achieve with unconstrained handwritten text.

3.3 Robustness of Classification Techniques Experiments have been carried to study the accuracy and effectiveness in writer identification using different classification techniques such as in [12]. In [12], WED and K-NN have been tested and the result shows that WED performed better than KNN with 96% accuracy as compared to 85% accuracy using same data set and features.

4. Conclusion Writer identification and verification have been researched over the past years aiming to add new discoveries to the field especially in terms of features that can uniquely identify writers and classification techniques that keep the phase simple yet achieve high accuracy. The research would continue as the demand for such applications are increasing along with the technology advances and problems that occur in the field that need to be solved. Results of developed systems showed great accuracy. However, there is no standard database which has been used to justify such results. In terms of features extracted, each feature has its own weaknesses and strength. For example, graphemes codebook construction needs a lot of training samples from the writers to cover all the possible set of graphemes but it keeps the classification step simple. Some features however are not suitable to be used for writer identification as it does not uniquely identify the writer itself. There are many classification techniques that can be used for writer identification. Some techniques are simple and easy to implement. Some involves expensive computational steps. All these need in depth research to solve the problems of writer identification.

5. References [1]

[2]

[3]

[4]

[5]

A. Namboodiri, S.Gupta, “Text Independent Writer Identification from Online Handwriting”, Proc. of 10th Int’l Workshop on Frontiers in Handwriting Recognition (IWFHR 2006), 23-26 October, La Baule, France J Chapran, “Biometrics Writer Identification Based on the Interdependency between Static and Dynamic Features of Handwriting”, Proc. of 10th Int’l Workshop on Frontiers in Handwriting Recognition (IWFHR 2006), 2326 October, La Baule, France L. Schomaker, M. Bulacu and K. Franke, “Automatic Writer Identification Using Fragmented Connected-Component Contours”, Proc. of the Ninth Int’l Workshop on Frontiers of Handwriting Recognition (IWFHR’04), pp 185-190, 2004. M. Bulacu, L. Schomaker and L Vuurpijl, “Writer Identification Using Edge-Based Directional Features”, Proc of 7th Int. Conf. on Document Analysis and Recognition (ICDAR 2003), IEEE Press, 2003, pp. 937-941, vol. II, 3-6 August, Edinburgh, Scotland. A. Bensefia, T. Paquet and L. Heutte, “Writer Identification By Writer’s Invariants”, Proc., 8th Int’l Workshop on Frontiers in Handwriting Recognition, IWFHR’02, Niagara-on-the-Lake, Canada, 6-8 Aug. 2002, pp 274-279

[6]

[8]

[9]

[10]

[11]

[12]

[13]

[14] [15]

[16]

A. Bensefia, T. Paquet and L. Heutte, “Handwritten Document Analysis for Automatic Writer Recognition”, Electronic Letters on Computer Vision and Image Analysis 5(2):72-86, 2005. A. Schapbach, B. Bunke, “Using HMM Based Recognizers for Writer Identification and Verification”, Proc. of the 9th Int’l Workshop on Frontiers of Handwriting Recognition (IWFHR-9 2004), 26-29 Oct 2004, pp 167-172. E. N. Zois, V, Anastassopoulos, “Morphological Waveform Coding for Writer Identification”, Pattern Recognition 33 (2000), pp. 385-386. M. Bulacu, L.Schomaker, “Combining Multiple Features for Text-Independent Writer Identification and Verification”, Proc. of 10th Int’l Workshop on Frontiers in Handwriting Recognition (IWFHR 2006), 23-26 October, La Baule, France. Laurens van der Maaten, “Improving Automatic Writer Identification”, Universiteit Maastricht, Institute for Knowledge and Agent Technology, The Netherlands, August 2005 H. E. S. Said, T. N. Tan, K. D. Baker, “Personal Identification based on Handwriting”, Pattern Recognition 33 (2000), 149-160, 1999 Pattern Recognition Society. Sargur N. Srihari, “Handwriting Identification: Research to Study Validity of Individuality of Handwriting & Develop Computer-Assisted Procedures for Comparing Handwriting”, State University of New York. Buffalo, September 2000. A. Webb, “Statistical Pattern Recognition”, Second Edition, John Wiley & Sons Ltd, 2002. L. Shomaker, M. Bulacu and M. van Erp, “Sparse-parametric Writer Identification using Heterogeneous Feature Groups”, Proc. of Int. Conf. on Image Processing (ICIP 2003), IEEE Press, 2003, pp 545-548, vol. I, 14-17 September, Barcelona, Spain F. Shahabi and M. Rahman, “Comparison of Gabor-based Features for Writer Identification of Farsi / Arabic Handwriting”, Proc. of 10th Int’l Workshop on Frontiers in Handwriting Recognition (IWFHR 2006), 23-26 October, La Baule, France

Writer Identification and Verification: A Review - Semantic Scholar

in the database. Most of the present ... reference database in the identification process. From these two ..... Heterogeneous Feature Groups”, Proc. of Int. Conf. on ...

138KB Sizes 0 Downloads 279 Views

Recommend Documents

Writer Identification and Verification: A Review - Semantic Scholar
Faculty of Information & Communication Technology ... verification: feature extraction phase, classification phase ... cons of each of the writer identification systems. .... A stroke ending is defined as ..... Handwriting & Develop Computer-Assisted

Text-Independent Writer Identification and Verification ...
writer identification and verification performance in exten- sive tests carried out using large datasets (containing up to. 900 subjects) of Western handwriting [3].

Text-Independent Writer Identification and Verification ...
it is necessary to use computer representations (features) with the ability to ... in a handwriting database with the return of a likely list of candidates) and writer ...

Efficient Speaker Identification and Retrieval - Semantic Scholar
Department of Computer Science, Bar-Ilan University, Israel. 2. School of Electrical .... computed using the top-N speedup technique [3] (N=5) and divided by the ...

Efficient Speaker Identification and Retrieval - Semantic Scholar
identification framework and for efficient speaker retrieval. In ..... Phase two: rescoring using GMM-simulation (top-1). 0.05. 0.1. 0.2. 0.5. 1. 2. 5. 10. 20. 40. 2. 5. 10.

Identification of Parametric Underspread Linear ... - Semantic Scholar
Feb 5, 2011 - W.U. Bajwa is with the Department of Electrical and Computer Engineering, ... 1. Schematic representation of identification of a time-varying linear ..... number of temporal degrees of freedom available for estimating H [8]: N ...... bi

Extension of Linear Channels Identification ... - Semantic Scholar
1Department of Physics, Faculty of Sciences and Technology, Sultan Moulay ... the general case of the non linear quadratic systems identification. ..... eters h(i, i) and without any information of the input selective channel. ..... Phase (degrees).

Identification of Parametric Underspread Linear ... - Semantic Scholar
Feb 5, 2011 - converter; see Fig. 2 for a schematic ... as the Kτ -length vector whose ith element is given by Ai (ejωT ), the DTFT of ai[n]. It can be shown.

Tansley review - Semantic Scholar
Jan 15, 2006 - discussion by highlighting new tools, developments and foci that will probably ...... analysis tools to functional genes related to soil aggregation,.

review articles - Semantic Scholar
Nov 2, 2014 - background, religion, and other life experi- ences. ... development of the implicit asso- ciation test and ... In software development, we see this.

Tansley review - Semantic Scholar
Jan 15, 2006 - eds. Arbuscular mycorrhizas physiology and function. Dordrecht, the ..... We are committed to rapid processing, from online submission through ...

Towards Explainable Writer Verification and ...
fication and identification systems have been implemented and impressive ... These systems usu- ..... systems usually are into a more transparent one, since ba-.

Person Identification based on Palm and Hand ... - Semantic Scholar
using Pieas hand database is 96.4%. 1. ... The images in this database are captured using a simple .... Each feature is normalized before matching score to.

Identification of potential maintainers and restorers ... - Semantic Scholar
under bagged condition. Potential maintainers were identified as having >90% pollen sterility and

speaker identification and verification using eigenvoices
approach, in which client and test speaker models are confined to a low-dimensional linear ... 100 client speakers for a high-security application, 60 seconds or more of ..... the development of more robust eigenspace training techniques. 5.

Person Identification based on Palm and Hand ... - Semantic Scholar
amount of variance among the images and the last dimension of this subspace ... A covariance matrix is created by multiplying the data matrix with its transpose.

A Comparison of Clustering Methods for Writer Identification and ...
a likely list of candidates. This list is ... (ICDAR 2005), IEEE Computer Society, 2005, pp. 1275-1279 ... lected from 250 Dutch subjects, predominantly stu- dents ...

Online Signature Verification using Dynamic Time ... - Semantic Scholar
Online Signature Verification, Dynamic Time Warping ... applications such as bank checks. ... achieved the best result with Equal Error Rate of 1.46%. H. Lei et al ...

A Appendix - Semantic Scholar
buyer during the learning and exploit phase of the LEAP algorithm, respectively. We have. S2. T. X t=T↵+1 γt1 = γT↵. T T↵. 1. X t=0 γt = γT↵. 1 γ. (1. γT T↵ ) . (7). Indeed, this an upper bound on the total surplus any buyer can hope

A Appendix - Semantic Scholar
The kernelized LEAP algorithm is given below. Algorithm 2 Kernelized LEAP algorithm. • Let K(·, ·) be a PDS function s.t. 8x : |K(x, x)| 1, 0 ↵ 1, T↵ = d↵Te,.

An Empirical Study on Uncertainty Identification in ... - Semantic Scholar
etc.) as information source to produce or derive interpretations based on them. However, existing uncertainty cues are in- effective in social media context because of its specific characteristics. In this pa- per, we propose a .... ity4 which shares

Identification of putative trait based markers for ... - Semantic Scholar
low productivity of Eucalyptus plantations in India, ... at http://dendrome.ucdavis.edu/ index.php. In eucalypts ... The present paper highlights the development of.

identification of pacemaking region in zebrafish ... - Semantic Scholar
In previous work, voltage dynamics have been visualized and activation maps have been manually drawn based on the visualization [8, 29, 47]. Manual methods are ... How do we perform computer assisted manual identification on optical mapping data? ...

Efficient Tag Identification in Mobile RFID Systems - Semantic Scholar
model, we propose efficient solutions to identify moving RFID tags, according to the .... the amount of energy reflected by an object is dependent on the reflective area ...... an alternative scheme to estimate the number of tags. Our methodology ...