The current state-of-the-art, in terms of performance, for solving document image binarization is training artificial neural networks on pre-labelled ground truth data. As such, it faces the same issues as other, more conventional, classification problems; requiring a large amount of training data. However, unlike those conventional classification problems, document image binarization involves having to either manually craft or estimate the binarized ground truth data, which can be error-prone and time-consuming. This is where sample selection, the act of selecting training samples based on some method or metric, might help. By reducing the size of the training dataset in such a way that the binarization performance is not impacted, the required time spent creating the ground truth is also reduced. This paper proposes a cluster-based sample selection method that uses image similarity metrics and the relative neighbourhood graph to reduce the underlying redundancy of the dataset. The method, implemented with affinity propagation and the structural similarity index, reduces the training dataset on average by 49.57% while reducing the binarization performance only by 0.55%.

Over the last decades companies and government institutions have gathered vast collections of images of historical handwritten documents. In order to make these collections truly useful to the broader public, images suffering from degradations, such as faded ink, bleed through or stains, need to be made readable and the collections as a whole need to be made searchable. Readability can be achieved by separating text foreground from page background using document image binarization, while searchability by search string or by example image can be achieved through word spotting. Developing algorithms with reasonable binarization or word spotting performance is a difficult task. Additional challenges are to make these algorithms execute fast enough to process vast collections of images in a reasonable amount of time, and to enable them to learn from few labeled training samples. In this thesis, we explore heterogeneous computing, parameter prediction, and enhanced throughput as ways to reduce the execution time of document image binarization algorithms. We find that parameter prediction and mapping a heuristics based binarization algorithm to the GPU lead to an 1.7 and 3.5 increase in execution performance respectively. Furthermore, we identify for a learning based binarization algorithm using recurrent neural networks the number of pixels processed at once as way to trade off execution time with binarization quality. The achieved increase in throughput results in a 3.8 times faster overall execution time. Additionally, we explore guided machine learning (gML) as a possible approach to reduce the required amount of training data for learning based algorithms for binarization, character recognition and word spotting. We propose an initial gML system for binarization, which allows a user to improve an algorithm’s binarization quality by selecting suitable training samples. Based on this system, we identify and pursue three different directions, viz., formulation of a clear definition of gML, identification of an efficient knowledge transfer mechanism from user to learner, and automation of sample selection. We explore the Learning Using Privileged Information paradigm as a possible knowledge transfer mechanism by using character graphs as privileged information for training a neural network based character recognizer. Furthermore, we show that, given a suitable word image representation, automatic sample selection can help to reduce the amount of training data required for word spotting by up to 69%.

Large collections of historical document images have been collected by companies and government institutions for decades. More recently, these collections have been made available to a larger public via the Internet. However, to make accessing them truly useful, the contained images need to be made readable and searchable. One step in that direction is document image binarization, the separation of text foreground from page background. This separation makes the text shown in the document images easier to process by humans and other image processing algorithms alike. While reasonably well working binarization algorithms exist, it is not sufficient to just being able to perform the separation of foreground and background well. This separation also has to be achieved in an efficient manner, in terms of execution time, but also in terms of training data used by machine learning based methods. This is necessary to make binarization not only theoretically possible, but also practically viable.

In this thesis, we explore different ways to achieve efficient binarization in terms of execution time by improving the implementation and the algorithm of a state-of-the-art binarization method. We find that parameter prediction, as well as mapping the algorithm onto the graphics processing unit (GPU) help to improve its execution performance. Furthermore, we propose a binarization algorithm based on recurrent neural networks and evaluate the choice of its design parameters with respect to their impact on execution time and binarization quality. Here, we identify a trade-off between binarization quality and execution performance based on the algorithm’s footprint size and show that dynamically weighted training loss tends to improve the binarization quality. Lastly, we address the problem of training data efficiency by evaluating the use of interactive machine learning for reducing the required amount of training data for our recurrent neural network based method. We show that user feedback can help to achieve better binarization quality with less training data and that visualized uncertainty helps to guide users to give more relevant feedback.

With the growth of virtualization and cloud computing, more and more forensic investigations rely on being able to perform live forensics on a virtual machine using virtual machine introspection (VMI). Inspecting a virtual machine through its hypervisor enables investigation without risking contamination of the evidence, crashing the computer, etc. To further access to these techniques for the investigator/researcher we have developed a new VMI monitoring language. This language is based on a review of the most commonly used VMI-techniques to date, and it enables the user to monitor the virtual machine's memory, events and data streams. A prototype implementation of our monitoring system was implemented in KVM, though implementation on any hypervisor that uses the common x86 virtualization hardware assistance support should be straightforward. Our prototype outperforms the proprietary VMWare VProbes in many cases, with a maximum performance loss of 18% for a realistic test case, which we consider acceptable. Our implementation is freely available under a liberal software distribution license.

In the context of historical document analysis, image binarization is a first important step, which separates foreground from background, despite common image degradations, such as faded ink, stains, or bleed-through. Fast binarization has great significance when analyzing vast archives of document images, since even small inefficiencies can quickly accumulate to years of wasted execution time. Therefore, efficient binarization is especially relevant to companies and government institutions, who want to analyze their large collections of document images. The main challenge with this is to speed up the execution performance without affecting the binarization performance. We modify a state-of-the-art binarization algorithm and achieve on average a 3.5 times faster execution performance by correctly mapping this algorithm to a heterogeneous platform, consisting of a CPU and a GPU. Our proposed parameter tuning algorithm additionally improves the execution time for parameter tuning by a factor of 1.7, compared to previous parameter tuning algorithms. We see that for the chosen algorithm, machine learning-based parameter tuning improves the execution performance more than heterogeneous computing, when comparing absolute execution times. © 2018 The Author(s)

This paper compares three different word image representations as base for label free sample selection for word spotting in historical handwritten documents. These representations are a temporal pyramid representation based on pixel counts, a graph based representation, and a pyramidal histogram of characters (PHOC) representation predicted by a PHOCNet trained on synthetic data. We show that the PHOC representation can help to reduce the amount of required training samples by up to 69% depending on the dataset, if it is learned iteratively in an active learning like fashion. While this works for larger datasets containing about 1 700 images, for smaller datasets with 100 images, we find that the temporal pyramid and the graph representation perform better.

A major challenge in document image binarization is the large variety in appearance of images from different document collections. This is especially challenging for parameterless, machine learning based binarization algorithms, which require additional ground truth training data to generalize or fine-tune to a new image collection. Reducing this costly labeling effort is relevant to companies and government institutions, which possess many different document image collections. One approach to address this problem is interactive machine learning, which enables a user to guide the fine-tuning process by providing feedback on the produced binarization result.

In this paper, we evaluate the claim that user guided training requires less labeled samples to fine-tune a basic model for binarization to a new image collection. Further, we propose a way to guide user feedback by visualizing the model’s labeling uncertainty and analyze the relationship between model uncertainty and binarization quality. Our experiments show that user feedback biases the model towards favoring foreground labels, which results in less erased text and thus better readability than when training samples are chosen randomly. Additionally, we find that model uncertainty serves as a useful guide for users and explain how the Dunning-Kruger effect prevents model uncertainty from being useful for automated sample selection.

In a child’s development, the child’s inherent ability to construct knowledge from new information is as important as explicit instructional guidance. Similarly, mechanisms to produce suitable learning representations, which can be trans- ferred and allow integration of new information are important for artificial learning systems. However, equally important are modes of instructional guidance, which allow the system to learn efficiently. Thus, the challenge for efficient learning is to identify suitable guidance strategies together with suitable learning mechanisms.

In this paper, we propose guided machine learning as source for suitable guidance strategies, we distinguish be- tween sample selection based and privileged information based strategies and evaluate three sample selection based strategies on a simple transfer learning task. The evaluated strategies are random sample selection, i.e., supervised learning, user based sample selection based on readability, and user based sample selection based on readability and uncertainty. We show that sampling based on readability and uncertainty tends to produce better learning results than the other two strategies. Furthermore, we evaluate the use of the learner’s uncertainty for self directed learning and find that effects similar to the Dunning-Kruger effect prevent this use case. The learning task in this study is document image binarization, i.e., the separation of text foreground from page background and the source domain of the transfer are texts written on paper in Latin characters, while the target domain are texts written on palm leaves in Balinese script.

Involving humans in the learning process of a machine learning algorithm can have many advantages ranging from establishing trust into a particular model to added personalization capabilities to reducing labeling efforts. While these approaches are commonly summarized under the term interactive machine learning (iML), no unambiguous definition of iML exists to clearly define this area of research. In this position paper, we discuss the shortcomings of current definitions of iML and propose and define the term guided machine learning (gML) as an alternative.

In the context of document image analysis, image binarization is an important preprocessing step for other document analysis algorithms, but also relevant on its own by improving the readability of images of historical documents. While historical document image binarization is challenging due to common image degradations, such as bleedthrough, faded ink or stains, achieving good binarization performance in a timely manner is a worthwhile goal to facilitate efficient information extraction from historical documents. In this paper, we propose a recurrent neural network based algorithm using Grid Long Short-Term Memory cells for image binarization, as well as a pseudo F-Measure based weighted loss function. We evaluate the binarization and execution performance of our algorithm for different choices of footprint size, scale factor and loss function. Our experiments show a significant trade-off between binarization time and quality for different footprint sizes. However, we see no statistically significant difference when using different scale factors and only limited differences for different loss functions. Lastly, we compare the binarization performance of our approach with the best performing algorithm in the 2016 handwritten document image binarization contest and show that both algorithms perform equally well.

This paper proposes a pre-training method for neural network-based character recognizers to reduce the required amount of training data, and thus the human labeling effort. The proposed method transfers knowledge about the similarities between graph representations of characters to the recognizer by training to predict the graph edit distance. We show that convolutional neural networks trained with this method outperform traditional supervised learning if only ten or less labeled images per class are available. Furthermore, we show that our approach performs up to 33% better than a graph edit distance based recognition approach, even if only one labeled image per class is available. © 2019 IEEE.