Marc-Peter Schambach

Model length adaptation of an HMM based cursive word recognition system

On the basis of a well accepted, HMM-based cursive script recognition system, an algorithm which automatically adapts the length of the models representing the letter writing variants is proposed. An average improvement in recognition performance of about 2.72 percent could be obtained. Two initialization methods for the algorithm have been tested, which show quite different behaviors; both prove to be useful in different application areas. To get a deeper insight into the functioning of the algorithm a method for the visualization of letter HMMs is developed. It shows the plausibility of most results, but also the limitations of the proposed method. However, these are mostly due to given restrictions of the training and recognition method of the underlying system.

Fast script word recognition with very large vocabulary

For an HMM-based script word recognition system an algorithm for fast processing of large lexica is presented. It consists of two steps: First, a lexicon-free recognition is performed, followed by a tree search on the intermediate results of the first step, the trellis of probabilities. Thus, the computational effort for recognition itself can be reduced in the first step, while preserving recognition accuracy by the use of detailed information in the second step. A speedup factor of up to 15/spl times/ could be obtained compared to traditional tree recognition, making script word recognition with large lexica available to time-critical tasks like in postal automation. There, lexica with e.g. all city or street names (20-500 k) have to be processed within a few milliseconds.

Low resolution Arabic recognition with multidimensional recurrent neural networks

OCR of multi-font Arabic text is difficult due to large variations in character shapes from one font to another. It becomes even more challenging if the text is rendered at very low resolution. This paper describes a multi-font, low resolution, and open vocabulary OCR system based on a multidimensional recurrent neural network architecture. For this work, we have developed various systems, trained for single-font/single-size, single-font/multi-size, and multi-font/multi-size data of the well known Arabic printed text image database (APTI). The evaluation tasks from the second Arabic text recognition competition, organized in conjunction with ICDAR 2013, have been adopted. Ten Arabic fonts in six font size categories are used for evaluation. Results show that the proposed method performs very well on the task of printed Arabic text recognition even for very low resolution and small font size images. Overall, the system yields above 99% recognition accuracy at character and word level for most of the printed Arabic fonts.

Determination of the number of writing variants with an HMM based cursive word recognition system

An important parameter for building a cursive script model is the number of different, relevant letter writing variants. An algorithm performing this task automatically by optimizing the number of letter models in an HMM-based script recognition system is presented. The algorithm iteratively modified selected letter models; for selection, quality measures like HMM distance and emission weight entropy are developed, and their correlation with recognition performance is shown. Theoretical measures for the selection of overall model complexity are presented, but best results are obtained by direct selection criteria: likelihood and recognition rate of training data. With the optimized models, an average improvement in recognition rate of up to 5.8 percent could be achieved.

Stabilize Sequence Learning with Recurrent Neural Networks by Forced Alignment

Cursive handwriting recognition is still a hot topic of research, especially for non-Latin scripts. One of the techniques which yields best recognition results is based on recurrent neural networks: with neurons modeled by long short-term memory (LSTM) cells, and alignment of label sequence to output sequence performed by a connectionist temporal classification (CTC) layer. However, network training is time consuming, unstable, and tends to over-adaptation. One of the reasons is the bootstrap process, which aligns the label data more or less randomly in early training iterations. This also leads to the fact that the emission peak positions within a character are located unpredictably. But positions near the center of a character are more desirable: In theory, they better model the properties of a character. The solution presented here is to guide the back-propagation training in early iterations: Character alignment is enforced by replacing the forward-backward alignment by fixed character positions: either pre-segmented, or equally distributed. After a number of guided iterations, training may be continued by standard dynamic alignment. A series of experiments is performed to answer some of these questions: Can peak positions be controlled in the long run? Can training iterations be reduced, getting results faster? Is training more stable? And finally: Do defined character position lead to better recognition performance?

How Postal Address Readers Are Made Adaptive

In the following chapter we describe how a postal address reader is made adaptive. A postal address reader is a huge application, so we concentrate on technologies used to adapt it to a few important tasks. In particular, we describe adaptation strategies for the detectors and classifiers of regions of interest (ROI), for the classifiers for single character recognition, for a hidden Markov recogniser for hand written words and for the address dictionary of the reader. The described techniques have been deployed in all postal address reading applications, including parcel, flat, letter and in-house mail sorting.

A new view of the output from word recognition

Word recognition has changed in the past years. Implementations have become better, which allows larger vocabularies to be recognized; many of the implementations now available are suited well to specific tasks, but several have to be combined to obtain maximum benefit; and they are still hampered by deficiencies, for example when trying to decide among similar words. The new view of the output interface from word recognition presented here aims at various goals: easier combination of comparable word classification systems, better post-processing of word recognition results at higher levels (where more context is available), and improvement of out-of-vocabulary behavior. The proposed solution is to provide two different result scores instead of one single value: an overall quality measure, indicating the credibility of recognition, and a similarity measure for each word alternative. This shifts the responsibility for decisions from low-level word recognition tasks to higher levels, while retaining all necessary, recognition information. This, however, gives rise to the need to define new performance metrics for evaluation. Implementations of the proposed output interface for two recognition engines are sketched.

Reviewing Performance Metrics for Handwriting Recognition: Must-Rejects and Recognition Graph Scores

The performance of handwriting recognition systems is typically measured in terms of “recognition rate”. Many academic competitions work this way. However, additional external requirements may shift the view of recognition quality: Processing time and acceptable error rate may be limited, lexica may be missing, but are needed to unambiguously define result correctness. These aspects will be discussed in detail, and appropriate metrics will be proposed. A single-valued combination of these metrics may then be defined for specific application areas. It can be used in order to choose between recognition approaches or systems, and to optimize system parameters automatically.

Increasing Robustness of Handwriting Recognition Using Character N-Gram Decoding on Large Lexica

Offline handwriting recognition systems often include a decoding step, that is retrieving the most likely character sequence from the underlying machine learning algorithm. Decoding is sensitive to ranges of weakly predicted characters, caused e.g. by obstructions in the scanned document. We present a new algorithm for robust decoding of handwriting recognizer outputs using character n-grams. Multidimensional hierarchical subsampling artificial neural networks with Long-Short-Term-Memory cells have been successfully applied to offline handwriting recognition. Output activations from such networks, trained with Connectionist Temporal Classification, can be decoded with several different algorithms in order to retrieve the most likely literal string that it represents. We present a new algorithm for decoding the network output while restricting the possible strings to a large lexicon. The index used for this work is an n-gram index with tri-grams used for experimental comparisons. N-grams are extracted from the network output using a backtracking algorithm and each n-gram assigned a mean probability. The decoding result is obtained by intersecting the n-gram hit lists while calculating the total probability for each matched lexicon entry. We conclude with an experimental comparison of different decoding algorithms on a large lexicon.