Sij+(d) The result obtained for the d-th attribute of the j-th feature vector assigned to the i-th code vector/atom. Sk| The number of eigenvectors of all submoves segmented from the training samples of the k-th writer.
Introduction
Overview of writer identification systems
In the literature on handwriting analysis, the processing of such data is called "online". Furthermore, the vast majority of works in the literature on writer identification refer to the offline environment.
Previous works on online writer identification
Regarding the proposed methodology, first the writer's accent is predicted. Recently, deep learning-based approaches have recently attracted the interest of the online research community for writer identification.
Contributions of the thesis
- Chapter 2: Exploration of codebook descriptors
- Chapter 3: Exploration of sparse coding based descriptors
- Chapter 4: Exploration of saliency information in the sparse framework . 14
- A block schematic of our proposal
We begin by examining the merits of the VLAD approach to online author identification. Based on a distance measure, each feature vector of the document is assigned to a specific code vector in the codebook.
Preprocessing
To avoid such values, we consider the x and y coordinates among sets of points with the same timestamp and average their values. In such cases, we consider the time difference between the last point of a stroke and the first point of the subsequent stroke.
Feature Extraction
Codebook description
VLAD
In the VLAD formulation, each of the components fij(d)−μi(d), 1≤d≤D in the residual vector fij−μ contributes equally to the aggregation process - with both high and low deformations given equal preference (see equation 2.3). . As a demonstration of the above disadvantage, we consider in Figure 2.2 a toy example diagram of a Voronoi cell. 5For the current discussion, due to the nonlinear separability of writers 1 and 2, we focus on improving their discrimination regarding the identification problem.
Proposed Descriptor
In our work, the normalized scores in Equation 2.5 for each of the Dfeatures are concatenated as the descriptor Si for code vector µi. Due to this restriction, the document can be represented by contiguous descriptors corresponding to any of the M −1 code vectors. 1≤i≤M, 1≤j ≤ni, 1≤d≤D However, a better separation can be obtained when only one of the point functions Sij−(d) and Sij+(d) is chosen forfij(d), depending on the sign of the deformation fij(d)−µi(d).
Computational Complexity
For this illustration, we consider using a scoring function of the form of Equation 2.8 for all attributes - regardless of their distortion sign for the assigned code vector. Calculating the scores Sij+(d) and Sij−(d) in Equation 2.4 for each feature attribute and code vector for the proposed methodology results in a time complexity of O(D× max. 1≤i≤Mni). After that, the calculation of the descriptor Si for the code vector µi requires a complexity of O(D× max. 1≤i≤Mni) or O(D) for VLAD or the proposed methodology.
Writer Identification
Here/stands for "Not applicable", indicating that the corresponding operation is not relevant for VLAD. A similar comment applies to the calculation of ˜Si+(d) and ˜Si−(d) scores in equation 2.5. In the following we give a brief description of the same in the general pattern recognition setting.
Experimental set-up
Database Description
Training and testing protocol
Performance Evaluation
- Comparison with VLAD for varying codebook size M
- Influence of the gap parameter
- Comparison to variants of VLAD
- Empirical study with a reduced version of our descriptor
- Performance with a variant of writer descriptor
- Statistical Significance
However, as the values of r increase, the writer identification performance starts to decline - as much as it can. For completeness, in table 2.4 we reflect the obtained skewness values in relation to the histograms of Figure 2.7 and 2.8. The evaluation of the above descriptor at the paragraph and text line levels for the IAM and IBM-UB1 databases are shown in Table 2.8 (a) and (b).
Conclusion
Our idea focuses on capturing the similarity of the properties1 of each individual subtype with the corresponding value in the subset of dictionary atoms that contribute with a non-zero sparse coding coefficient. To keep this in perspective, we derive descriptions of each of the dictionary atoms, which include the similarity scores of the features in a feature vector. The features calculated for each substroke of the online trace are stacked to form a feature vector.
Proposed Methodology
Sub-stroke generation
Feature extraction
Entropy based selection of bin size B
At the first level, feature vectors corresponding to the embezzlements are collected among the set of sections of the W-writers and partitioned into k1 clusters. As an extension of the above, when each of the dek1×W subclusters generated in the second level includes feature vectors corresponding to only one writer, the average entropy measure is defined by . The entropy corresponding to the distribution of the authors in each of the three subclusters is calculated separately and then accumulated.
Sparse coding: an overview
The element (i, j) of this matrix contains αij, the sparse coefficient from the j. embezzlement corresponding to the idea dictionary's atomφφφi. The average pooling strategy to generate the writer descriptors from the sparse coefficients can be defined by a vector z of size M whose ide element is given as . In the following section, we propose descriptors that aim to capture additional information from the sparse coefficients, which somehow outperform the traditional max/average pooled strategies.
Proposed sparse coding based writer description
Discussion
As a first step in calculating the descriptor distance between writers, we consider the average of the distance between the city and the block obtained between the descriptors of two different writers w1, w2, as follows. The value of R is somehow related to the degree of separation of the writer's descriptors in the feature space. In Table 3.1, we show the results of the recorder data from the databases, namely IAM and IBM-UB1.
Results and Discussion
- Analysis of average entropy values H B with bin size B
- Performance with varying bins B and dictionary size M
- Influence of histogram feature sets on writer description
- Influence of the segmentation strategy
- Comparison with max and average pooling based descriptors
- Impact of the sparse framework for writer description
- Performance with a variant of writer descriptor
- Time complexity comparison with our proposal in Chapter 2
It is interesting to note that the value of the bin size B that leads to the best average author identification rate coincides with that which provides the lowest average entropy measure HB in Tables 3.3 and 3.4 for the IAM and IBM-UB1 databases. The evaluation of the above descriptor at the paragraph and text line levels for the IAM and IBM-UB1 databases is presented in Table 3.9. This can be attributed to the addition of the substroke generation module for the proposal in Chapter 3. b).
Conclusion
The number of histograms considered for this strategy corresponds to the size of the dictionary. In the following sections, we present the details of each of the contributions made in this Chapter. To begin, we provide a block diagram of our proposed framework that uses the salient information of dictionary atoms in Section 4.2.
Schematic of the proposed framework
Entropy based saliency computation
Histogram generation
The size of the containers is set to W - corresponding to the number of writers registered in the system. For each feature vector in the subset of the kth author (in subset Sk), we consider the sparse coefficient associated with the ith dictionary atom. Let {αkij} denote the sparse coefficient corresponding to the ith dictionary atom obtained from the jth embezzlement by the kth author (in Sk).
Computation of saliency with entropy based values
Likewise, the subfigures (c) and (d) are the histograms of the atoms presenting the second minimum and maximum value. The normalized votes for each of the four histograms are obtained by accumulating the sparse coefficients from the sub-lines corresponding to four sections of 50 authors of the IAM-Online database. The normalized votes for each of the four histograms are obtained by accumulating the sparse coefficients from the embezzlements corresponding to the data from the 43 authors of this database.
Sum-pooling based saliency computation
For each feature vector of the substroke of the k-th writer (in subset Sk), we consider the normalized sparse coefficient associated with the i-th dictionary atom. After generating the histogramH, the vote for the ith bin hi is given by. where each of the summation terms. corresponds to the normalized value of the sparse coefficient αkij which lies between [0,1]. To reduce the dynamic range of the salience values, we use the.
Modified Writer Descriptor
In our proposal, we try to modify each of the above elements (obtained from one of the strategies) by incorporating the available prior knowledge in the form of salient values for dictionary atoms. Our modified descriptor formulation ensures that higher preference is given to sparse coefficients corresponding to dictionary atoms with higher importance/importance and vice versa. Last but not least, we provide a high-level summary of the proposed writer descriptor in pseudo-code form (refer to Algorithm 5).
Writer specific saliency value adaptation
Proposed Identification Framework
Before using the ensemble kth SVM, we modify this vector by including their fitted saliency values precomputed for the kth writer. Note that ˆzk corresponds to the modified descriptor that is fed into the ensemble SVM kth. Use the modified zk descriptor as input to the ensemble kth SVM.
Experiments and discussion
Influence of the saliency based incorporation
In contrast, the traditional SPF descriptors lag behind, achieving a better average identification rate of 98.41% again at M = 400. We now demonstrate that the results of the proposed handwriting descriptors with the inclusion of salient values for dictionary atoms are statistically significant when compared to traditional mean and maximum clustering classification frameworks. A trend of low values for p is observed empirically in each of the entries in the table - thus indicating the statistical significance of our proposition.
Influence of writer-specific adaptation of saliency values
The results listed in Table 4.2 (a) and (b) on both databases show a promising trend with the modified writer descriptor at both paragraph and text line levels across all dictionary sizes. It can be noted here that the value β = 0 relates to the performance of the EN-SL system. From Table 4.5, we can see that the highest author identification rates at the paragraph and text line level for the IAM database are 99.54% and 91.26% for M= 400, respectively.
Conclusion
Proposal of a new codebook-based description that alleviates the limitation of the VLAD and its variants. Proposal of new histogram-based features for shape description of the sub-strips segmented from the online trace. The descriptions provide the description of the author at a finer level, compared to the traditional maximum / average pooling strategies.
Discussion of prior works
To the best of our knowledge, the author identification systems proposed in use the IAM database for experimentation. The latent Dirichlet allocation concept used in [13] and the subtractive clustering-based approach [44] lead to paragraph-level writer identification rates of 93.39% and 96.30%, respectively. To date, only one survey [13] has been performed on the IBM-UB1 database, with a writer identification rate of 89.47% reported at the paragraph level.
Possible pointers for the future
Tan, “Text-independent online author identification based on hit probability distribution function,” in International Conference on Biometrics. Tan, "Text-Independent Online Writer Identification Based on Time Sequence and Shape Codes." in Document Analysis and Recognition, 2009. Vivek Venugopal and Suresh Sundaram, “Web author identification with encoding-based sparse descriptors,” IEEE Trans.