Keywords:

• dominant points
• consecutive vectors
• toothbrush shape
• distance metric
• smallest perpendicular distance

Q1: How usual calculation of distance is done?

Minor DPs are deleted in approximation. A minor DP is a DP where the perpendicular distance between the point and the straight line is minimum.

a a
 b

Here b is deleted when its distance to the line a-a is minimum.

The perpendicular distance is calculated using

[ d_i = \sqrt{\frac{((x_i - x_a) (y_b - y_a) - (y_i - y_a) (x_b -x_a))^2}{(x_a - x_b)^2 + (y_a - y_b)^2}} ]

for lines between points $p_a$ and $p_b$ and the point $p_i$.

Q2: What is a toothbrush shape?

It’s something like

aaaaaa
bbbbbbbbbbbbbbbbbbbbbb

Hence the toothbrush.

Though I don’t get why is this particularly important.

Q3: Which information is included in distance metric?

Angle acuteness is added to the information described above.

Q4: How the distance metric differs from others?

It includes angle acuteness in the metric and the more acute the angle, the less likely it’s removed from DP set.

Q5: What’s baseline for performance and how does this improve it?

As the number of DPs decrease, RMSE of the new metric decreases. For large number of DPs it doesn’t matter much. (So performance penalty may not pay off)

URL: http://arxiv.org/pdf/0704.1267.pdf

Keywords:

• page segmentation
• overlapping components
• image quality
• document complexity
• preprocessing
• projection based
• smearing based
• grouping based
• hough transform based
• repulsive attractive
• stochastic
• touching components

Q1: What are the most usable techniques for Ottoman divans?

Likforman-Sulem and Faure’s technique, which uses Gestalt criteria to associate text elements, might be of use. Feldbach and Tennies’ work, which was tested on Church Registers, may also be helpful. The Hough transform can be used. The Repulsive-Attractive method of Öztop et al. is also applicable. Stochastic methods by Tseng and Lee, which use a probabilistic Viterbi algorithm, can also be utilized.

Q2: How are touching components successfully delimited?

A touching component can be detected by its size. Subsequently, it should either be assigned to a lower or upper line, or be separated. Successful separation requires letter images or skeletons (which we lack).

Q3: How is the Hough transform used?

Centroids of the connected components (CCs) are used as units of the Hough transform. Line hypotheses are developed in the Hough domain and verified in the image domain.

Q4: What are the problems specific to non-Latin texts?

The baseline of Hebrew is at the upper part of the letters because of their box shape. Devanagari and similar scripts also have a headline on top of them. Diacritics and inter-letter shapes pose problems for Arabic.

Keywords:

• ridge
• printed text
• non-text segmentation
• gaussian-filter bank
• reading order

Q1: How is line skew determined?

There is a $\theta$ parameter in the Gaussian kernel which is used to produce ridges. This may be used in detecting the skew, but since it’s constant for an entire page, a varying line skew will probably decrease its performance.

Q2: How are non-text portions detected?

The paper does not include a description but cites “S. S. Bukhari, F. Shafait, and T. M. Breuel, ‘Improved document image segmentation algorithm using multiresolution morphology,’ in Proc. SPIE Document Recognition and Retrieval XVIII, San Jose, CA, USA, Jan. 2011” as a source for an improved technique.

Q3: Which heuristics are used in reading order determination?

Breuel is reported to have an algorithm in “T. M. Breuel, ‘High performance document layout analysis,’ in Symposium on Document Image Understanding Technology, Greenbelt, MD, USA, April 2003.” The paper says the authors modified the algorithm for right-to-left scripts. No further details are provided.

Q4: How large is the dataset, and what does it contain?

25 Arabic documents and 20 Urdu documents are used.

Q5: Are there any techniques applicable to divans?

There might be, if any of them were described in detail. We already have more sophisticated text line detection techniques. For the others, I’ll need to read the cited works.

Keywords:

• handwritten
• paleography
• fragments
• SIFT
• sparse coding
• dictionaries

Q1: What is the ultimate goal of the authors?

The two main goals are providing tools to bring together fragments of the same page (specifically from the Cairo Genizah) and trying to classify handwriting and dates.

Q2: How is SIFT used?

SIFT is used at various points of a letter to generate descriptors. There are 100,000 descriptors overall before inputting them into k-means. SIFT serves as the main classification technique.

Q3: How did they produce the letter dictionaries?

They produced letter dictionaries using the generated SIFT descriptors and k-means clustering to find representative visual words.

Q4: What is sparse coding and its importance?

Sparse coding is used to code documents (or any other codable thing) with a separating/descriptive code which also shows the similarity between items. An example might be the bag of visual words approach.

Q5: Are there any relevant techniques for our research?

This is relevant to the Divan matching problem. This work might be cited in historical document matching, although it covers techniques that are generally well-known in the field.

Keywords:

• LMS estimator
• O(n logn)
• bracelet
• slab
• random
• approximation
• quantiles

Q1: What is LMS?

Given a set of points $p_0, …, p_n$, LMS finds a line $q_0, q_1$ that minimizes the median of the square of distances of $p_0, …, p_n$. This is in contrast with summing up all the squared distances and minimizing them as in OLS (Ordinary Least Squares).

Q2: How are approximations done using LMS?

There are exact solutions for the LMS problem. This paper presents an algorithm with lower complexity. It tries to find an LMS approximation in a band defined by a parameter $\epsilon_r$.

Q3: On which parameters does the algorithm depend?

The algorithm approxLMS depends on a set of lines, a set of quantiles, and two error bounds $\epsilon_r$ and $\epsilon_q$.

Q4: How can this help in our line approximations?

This paper is aimed towards providing an efficient algorithm for random approximations. Since we need repeatability in our keypoint detection, it might be harder (or impossible) to prove that the algorithm produces the exact same line endings in each run with a similar set of points. Hence, it’s not usable in our studies.