The ICDAR 2019 Time-Quality Binarization Competition assessed the performance of seventeen new together with thirty previously published binarization algorithms. The quality of the resulting two-tone image and the execution time were assessed. Comparisons were on both in "real-world" and synthetic scanned images, and in documents photographed with four models of widely used portable phones. Most of the submitted algorithms employed machine learning techniques and performed best on the most complex images. Traditional algorithms provided very good results at a fraction of the time.

This paper presents an offline handwriting recognition system for Bangla script using sequential detection of characters and diacritics with a Faster R-CNN. This is an entirely segmentation-free approach where the characters and associated diacritics are detected separately with different networks named C-Net and D-Net. Both of these networks were prepared with transfer learning from VGG-16. The essay scripts from the Boise State Bangla Handwriting Dataset along with standard data augmentation techniques were used for training and testing. The F1 scores for the C-Net and D-Net networks are 89.6% and 93.2% respectively. Afterwards, both of these detection modules were fused into a word recognition unit with CER (Character Error Rate) of 11.2% and WER (Word Error Rate) of 24.4%. A spell checker further minimized the errors to 8.9% and 21.5% respectively. This same method is likely to be equally effective on several other Abugida scripts similar to Bangla.

Type designers and historians studying the typefaces and fonts used in historical documents can usually only rely on available printed material. The initial wooden or metal cast fonts have mostly disappeared. In this paper we address the creation of character templates from printed documents. Images of characters scanned from Renaissance era documents are segmented, then clustered. A template is created from each obtained cluster of similar appearance characters. In order for subsequent typeface analysis tools to operate, the template should reduce the noise present in the individual instances by using information from the set of samples, but the samples must be homogeneous enough to not introduce further noise into the process. This paper evaluates the efficiency of several clustering algorithms and the associated parameters through cluster validity statistics and appearance of the resulting template image. Clustering algorithms that form tight clusters produce templates that highlight details, even though the number of available samples is smaller, while algorithms with larger clusters better capture the global shape of the characters.

Image binarization has a large effect on the rest of the document image analysis processes in character recognition. Algorithm development is still a major focus of research. Evaluation of image binarization has been done by comparison of the result of OCR systems on images binarized by different methods. That has been criticized in that the binarization alone is not evaluated, but rather how it interacts with the downstream processes. Recently pixel accurate "ground truth" images have been introduced for use in binarization algorithm evaluation. This has been shown to be open to interpretation. The choice of binarization ground truth affects the binarization algorithm design, either directly if design is by automated algorithm trying to match the provided ground truth, or indirectly if human designers adjust their designs to perform better on the provided data. Three variations in pixel accurate ground truth were used to train a binarization classifier. The performance can vary significantly depending on choice of ground truth, which can influence binarization design choices.

Vote counting accuracy has become a well-known issue in the vote collection process. Digital image processing techniques can be incorporated in the analysis of printed election ballots. Current image processing techniques in the vote collection process are heavily dependent on the anticipated, geometric positioning of the vote. These techniques don't account for markings made outside of the requested field of input. Using various form dropout techniques, however, every mark on the form can be extracted and used by the machine to make an intelligent decision. Most methods will still miss a few marks and result in a few false alarms. This paper explores methods of voting between the results of the different mark extraction methods to improve recognition. To provide diversity a simple image subtraction technique is paired with a distance transform and a morphology based algorithm. The result has a higher detection rate and a lower false alarm rate.

Resources are presented for fostering paper-based election technology. They comprise a diverse collection of real and simulated ballot and survey images, and software tools for ballot synthesis, registration, segmentation, and ground truthing. The grids underlying the designated location of voter marks are extracted from 13,315 degraded ballot images. The actual skew angles of sample ballots, recorded as part of complete ballot descriptions compiled with the interactive ground-truthing tool, are compared with their automatically extracted parameters. The average error is 0.1 degrees. These results provide a baseline for the application of digital image analysis to the scrutiny of electoral ballots.

The goal of this work is to add the capability to segment documents containing text, graphics, and pictures in the open source OCR engine OCRopus. To achieve this goal, OCRopus' RAST algorithm was improved to recognize non-text regions so that mixed content documents could be analyzed in addition to text-only documents. Also, a method for classifying text and non-text regions was developed and implemented for the Voronoi algorithm enabling users to perform OCR on documents processed by this method. Finally, both algorithms were modified to perform at a range of resolutions. Our testing showed an improvement of 15-40% for the RAST algorithm, giving it an average segmentation accuracy of about 80%. The Voronoi algorithm averaged around 70% accuracy on our test data. Depending on the particular layout and idiosyncracies of the documents to be digitized, however, either algorithm could be sufficiently accurate to be utilized.

Optical scan voting is considered by many to be the most trustworthy option for conducting elections because it provides an independently verifiable record of each voter’s intent. While op-scan technology has been in use for decades, attempts to improve the machine reading of ballots raises a range of interesting issues in document image analysis. Work thus far has been hindered by a lack of real-world data, since ballots associated with actual elections are kept secure from the public and normally destroyed after a period time. Fortunately, as a result of a recent challenged election in the State of Minnesota, a large collection of op-scan ballot images was made available for public inspection on the World Wide Web. In this paper, we present this unique resource to the document analysis community. We also describe our efforts to annotate the collection, including the latest version of a graphical tool we have developed for collecting ground-truth interpretations, along with the protocol now being employed. The collection, consisting of ballot images, file formats, and associated truth data, is being made openly available to facilitate research in this important area.

The accuracy of a binarization algorithm is often calculated relative to a ground truth image. Except for synthetically generated images, no ground truth image exists. Evaluating binarization on real images is preferred. The ground truthing between and among different operators is compared. Four direct metrics were used. The variability of the results of five different automatic binarization algorithms were compared to that of manual ground truth results. Significant variability in the ground truth results was found.

We compare in this study two image restoration approaches for the pre-processing of printed documents:namely the Non-local Means filter and a total variation minimization approach. We apply these two approaches to printed document sets from various periods,and we evaluate their effectiveness through character recognition performance using an open source OCR. Our results show that for each document set, one or both pre-processing methods improve character recog-nition accuracy over recognition without preprocessing. Higher accuracies are obtained with Non-local Means when characters have a low level of degradation since they can be restored by similar neighboring parts of non-degraded characters. The Total Variation approach is more effective when characters are highly degraded and can only be restored through modeling instead of using neighboring data.

Portable ballot counters using camera technology and manual paper feed are potentially more reliable and less expensive than scanner based systems. We show that the spatial sampling rate, geometric linearity, point spread function, and photometric transfer function of off-the-shelf consumer cameras are acceptable for ballot imaging. However, scanner illumination is much more uniform than can be economically accomplished for variable size ballots. Therefore flat-field compensation must be designed into the image processing software. We illustrate the mechanical design of a prototype camera based ballot reader based on our comparative observations.

The push toward voting via hand marked paper ballots has focused attention on the limitations of current optical scan systems. Discrepancies between human and machine interpretations of ballot markings can lead to a loss of trust in the election process. In this paper, a style-based approach to ballot recognition is proposed in which marks are recognized collectively rather than in isolation. The consistency of a voter's style is leveraged to improve the overall accuracy of the system. We compare style-based recognition to various kinds of singlet classifiers and show that it outperforms them by a substantial margin.

As a result of well-publicized security concerns with direct recording electronic (DRE) voting, there is a growing call for systems that employ some form of paper artifact to provide a verifiable physical record of a voter's choices. In this paper, we present a system we are developing to support a multi-institution, cross-disciplinary research project examining issues that arise when paper ballots are used in elections. We survey the motivating factors behind our work, discuss the special constraints raised in processing ballots as opposed to more general document images, and describe the current status of our system.

Because most degraded documents are created by people, the preferences individuals have in relation to degraded documents are quite important. Their preferences may determine whether or not the documents they created are appropriate for machines. The goal of this study was to find relationships between preference and several parame- ters of a scanner degradation model. It was found that the difference in binarization threshold and the difference in edge displacement caused by the degradation both had strong linear relationships to preference. The width of the point spread function did not show such a relationship. These relationships were counterintuitive because degraded characters with thicker stroke widths than the original were preferred to those that had stroke widths closer to the original character.

Printing and scanning of text documents introduces degradations to the characters which can be modeled. Interestingly, certain combinations of the parameters that govern the degradations introduced by the printing and scanning process affect characters in such a way that the degraded characters have a similar appearance, while other degradations leave the characters with an appearance that is very different. It is well known that (generally speaking), a test set that more closely matches a training set is recognized with higher accuracy than one that matches the training set less well. Likewise, classifiers tend to perform better on data sets that have lower variance. This paper explores an analytical method that uses a formal printer/scanner degradation model to identify the similarity between groups of degraded characters. This similarity is shown to improve the recognition accuracy of a classifier through model directed choice of training set data.

It is desirable to convert paper text documents to a computer readable and searchable form. For current technology, the combination of a scanner and Optical Character Recognition (OCR) software brings us closer to this goal. However, the degradations introduced to digital images during the scanning process significantly reduce the accuracy of OCR. Because of this imperfection, a degradation model has been developed to predict how a document image will look after being subjected to the appropriate scanning process. Methods exist to calibrate the model parameters using specialized charts such as large wedges. With a calibrated model, controlled experiments can be conducted to generate a large set of synthetic characters. These characters can then be used as the data in the training set for the OCR software, which should increase the accuracy rate of the classification process could increase. Character corners serve as a potential source to calibrate the degradation model. In this thesis, the acute corners in the sans-serif font text images kvwxyzAKMVWXYZ are used to characterize the degradation model. These characters are more readily available than large wedges. If characters’ corners can provide a high confidence level at estimating the parameters of the degradation model, they may allow us to calibrate a scanner without specialized calibration images. Many aspects of character corners are examined. The disadvantage of using character corners is that they are much smaller than large wedges. Misusing the information of the character corners increases the probability of providing a poor estimate. Synthetically generated large wedges are first investigated and used to estimate the degradation model to determine the constraints of the wedges. These results provide guidelines as to how character corners can be used robustly. They also allow us to judge the accuracy level of the estimation results from character corners. Large wedges theoretically estimate the parameters of the degradation model the best because their signal-to-noise ratio is high. The quantity of character corners is very limited on a typical page. Experiments using different resolutions of a limited quantity of synthetically generated characters to estimate the degradation model are conducted. The estimation results from highresolution (1200dpi) characters are better than from low-resolution (600dpi) characters. The high-resolution characters provide estimation results that are comparable to the large wedges. Character images on paper are scanned and used to estimate the degradation model. Large quantities of corners from characters are used to investigate how their contribution affects the mean and the standard deviation of the parameter estimators. The relationship between the angles of the corners used in estimation and the estimation results is studied. The number of typical pages of character images required to offer a reasonable result is also examined. Experimentation shows that using selected angles from characters improves the estimation results. Some characters contribute to estimators more than the others.

Scanning a high-contrast image in bilevel mode results in image degradation. This is caused by two primary effects: blurring and thresholding. This paper expands on a method of estimating a joint distortion parameter called the edge spread, from a star sector test chart in order to calculate the values of the point spread function width and binarization threshold. This theory is also described for variations in the source pattern which can represent degradations caused by repetition of the bilevel process as would be seen in printing then scanning, or in repeated photocopying. Estimation results are shown for the basic and extended cases.