Optical character recognition (OCR) software has advanced greatly in recent years. Machine-printed text can be scanned and converted to searchable text with word accuracy rates around 98%. Reasonably neat hand-printed text can be recognized with about 85% word accuracy. However, cursive handwriting still remains a challenge, with state-of-the-art performance still around 75%. Algorithms based on hidden Markov models have been only moderately successful, while recurrent neural networks have delivered the best results to date. This thesis explored the feasibility of using a special type of feedforward neural network to convert freeform cursive handwriting to searchable text. The hidden nodes in this network were grouped into clusters, with each cluster being trained to recognize a unique character bigram. The network was trained on writing samples that were pre-segmented and annotated. Post-processing was facilitated in part by using the network to identify overlapping bigrams that were then linked together to form words and sentences. With dictionary assisted post-processing, the network achieved word accuracy of 66.5% on a small, proprietary corpus. The contributions in this thesis are threefold: 1) the novel clustered architecture of the feed-forward neural network, 2) the development of an expanded set of observers combining image masks, modifiers, and feature characterizations, and 3) the use of overlapping bigrams as the textual working unit to assist in context analysis and reconstruction.

Various document layout analysis techniques are employed in order to enhance the accuracy of optical character recognition (OCR) in document images. Type-specific document layout analysis involves localizing and segmenting specific zones in an image so that they may be recognized by specialized OCR modules. Zones of interest include titles, headers/footers, paragraphs, images, mathematical expressions, chemical equations, musical notations, tables, circuit diagrams, among others. False positive/negative detections, oversegmentations, and undersegmentations made during the detection and segmentation stage will confuse a specialized OCR system and thus may result in garbled, incoherent output. In this work a mathematical expression detection and segmentation (MEDS) module is implemented and then thoroughly evaluated. The module is fully integrated with the open source OCR software, Tesseract, and is designed to function as a component of it. Evaluation is carried out on freely available public domain images so that future and existing techniques may be objectively compared. / Master of Science

Design and evaluation of input methods for secondary tasks in dual-task systems presents specific challenges not covered by traditional human-computer interaction design techniques. Emerging trends in the fields of mobile, ubiquitous, and in-vehicle information systems demonstrate a desire for users to interact with information systems while engaging in other tasks. Research on interaction within these various fields has revealed input methods that perform well for a particular task. However, few focus on the tradeoffs of attention that must be made to react to this notification information. A design analysis technique for input methods is proposed focusing on the design objectives of interruption, reaction, and comprehension for the secondary task made at the cost of primary task attention. Through a study conducted using a reusable usability test platform constructed for this thesis, a typical in-vehicle information system is analyzed using the proposed design analysis. Three input methods were designed and compared: a graffiti character recognizer, a touch screen, and a remote control for their proficiency at selecting an item from a list while operating a driving simulator. The results of the study revealed similar task performance between the varied input methods; however, the design analysis enabled recommendations about future design directions, confirming the viability of the technique for notification systems research. / Master of Science

By organically fusing the methods of convolutional neural network (CNN) and recurrent neural network (RNN), this dissertation focuses on the application of optical character recognition and image classification processing. The first part of this dissertation presents an end-to-end novel receipt recognition system for capturing effective information from receipts (CEIR). The main contributions of this research part are divided into three parts. First, this research develops a preprocessing method for receipt images. Second, the modified connectionist text proposal network is introduced to execute text detection. Third, the CEIR combines the convolutional recurrent neural network with the connectionist temporal classification with maximum entropy regularization as a loss function to update the weights in networks and extract the characters from receipt. The CEIR system is validated with the scanned receipts optical character recognition and information extraction (SROIE) database. Furthermore, the CEIR system has strong robustness and can be extended to a variety of different scenarios beyond receipts. For the convolutional recurrent structure application of land use image classification, this dissertation comes up with a novel deep learning model for land use classification, the convolutional recurrent land use classifier (CRLUC), which further improves the accuracy in classifying remote sensing land use images. Besides, the convolutional fully-connected neural networks with hard sample memory pool structure (CFMP) is invented to tackle the remote sensing land use image classification tasks. The CRLUC and CFMP algorithm performances are tested in popular datasets. Experimental studies show the proposed algorithms can classify images with higher accuracy and fewer training episodes compared to popular image classification algorithms.

Este trabalho apresenta um estudo sabre reconhecimento visual de caracteres através da utilização das redes neurais. São abordados os assuntos referentes ao Processamento Digital de Imagens, aos sistemas de reconhecimento de caracteres, e as redes neurais. Ao final é apresentada uma proposta de implementação de um sistema OCR orientado ao reconhecimento de caracteres impressos, que utiliza uma rede neural desenvolvida especificamente para esta aplicação. O sistema proposto, que é denominado de sistema N2OCR, possui um protótipo implementado que também é descrito neste trabalho. Em relação ao Processamento Digital de Imagens são apresentados diversos temas, abrangendo os assuntos referentes à aquisição de imagens, ao tratamento das imagens e ao reconhecimento de padrões. A respeito da aquisição de imagens são destacados os aspectos referentes aos dispositivos de aquisição e os tipos de imagens obtidas através destes. Sobre o tratamento de imagens são abordados os aspectos referentes a imagens textuais, incluindo: halftoning, geração e modificação de histograma, limiarização e operações de filtragem. Quanto ao reconhecimento de padrões é feita uma breve análise das técnicas relacionadas a este tema. Os diversos tipos de sistemas de reconhecimento de caracteres são abordados, assim coma as técnicas e algoritmos empregados por estes. Além destes tópicos é apresentada uma discussão a respeito da avaliação dos resultados obtidos por estes sistemas, assim como é feita uma análise das principais dificuldades enfrentadas por estas aplicações. Neste trabalho é feita uma apresentação a respeito das redes neurais, suas características, histórico e evolução das pesquisas nesta área. É feita uma descrição dos principais modelos de redes neurais em destaque na atualidade: Perceptron, Adaline, Madaline, redes multinível, ART, modelo de Hopfield, máquina de Boltzmann, BAM e modelo de Kohonen. A partir da análise dos diferentes modelos de redes neurais empregados na atualidade, chega-se a proposta de um novo modelo de rede a ser utilizado pelo sistema N2OCR. São descritos os itens referentes ao aprendizado, ao reconhecimento e as possíveis extensões deste novo modelo. Também é abordada a possibilidade de implementação de um hardware dedicado para este modelo. No final deste trabalho é fornecida uma visão global do sistema N2OCR, descrevendo cada um de seus módulos. Também é feita uma descrição do protótipo implementado e de suas funções. / This work presents a study of visual character recognition using neural networks. It describes some aspects related to Digital Image Processing, character recognition systems and neural networks. The implementation proposal of one OCR system, for printed character recognition, is also presented. This system uses one neural network specifically developed for this purpose. The OCR system, named N2OCR, has a prototype implementation, which is also described. Several topics related to Digital Image Processing are presented, including some referent to image acquisition, image processing and pattern recognition. Some aspects on image acquisiton are treated, like acquisition equipments and kinds of image data obtained from those equipments. The following items about text image processing are mentioned: halftoning, hystogram generation and alteration, thresholding and filtering operations. A brief analysis about pattern recognition related to this theme is done. Different kinds of character recognition systems are described, as the techniques and algorithms used by them. Besides, a di cussi on about performance estimation of this OCR systems is done, including typical OCR problems description and analysis. In this work, neural networks are presented, describing their characteristics, historical aspects and research evolution in this field. Different famous neural network models are described: Perceptron, Adaline, Madaline, multilevel networks. ART, Hopfield's model , Boltzmann machine, BAM and Kohonen's model. From the analysis of such different neural network models, we arrive to a proposal of a new neural net model, where are described items related to learning, recognition and possible model extensions. A possible hardware implementation of this model is also presented. A global vision of N2OCR system is presented at the end of this work, describing each of its modules. A description of the prototype implementation and functions is also provided.

No description available.

<p>DARC is a technique for data broadcasting over the FM radio network. Sectra Wireless Technologies AB has developed a handheld DARC receiver known as the Sectra CitySurfer. The CitySurfer is equipped with a high-resolution display along with buttons and a joystick that allows the user to view and navigate through various types of information received over DARC. </p><p>Sectra Wireless Technologies AB has, among other services, also developed a paging system that enables personal message transmission over DARC. The background of this thesis is a wish to be able to send fax documents using the paging system and to be able to view received fax documents in the CitySurfer. </p><p>The presented solution is a central PC-based fax server. The fax server is responsible for receiving standard fax transmissions and converting the fax documents before redirecting them to the right receiver in the DARC network. The topics discussed in this thesis are fax document routing, fax document conversion and fax server system design.</p>

Structured representation of input physical patterns as a set of local features has been useful for a veriety of robotics and human computer interaction (HCI) applications. It enables a stable understanding of the variable inputs. However, this representation does not fit the conventional machine learning algorithms and distance metrics because they assume vector inputs. To learn from input patterns with variable structure is thus challenging. To address this problem, I propose a general and systematic method to design distance metrics between structured inputs that can be used in conventional learning algorithms. Based on the observation of the stability in the geometric distributions of local features over the physical patterns across similar inputs, this is done combining the local similarities and the conformity of the geometric relationship between local features. The produced distance metrics, called “parametric kernels”, are positive semi-definite and require almost linear time to compute. To demonstrate the general applicability and the efficacy of this approach, I designed and applied parametric kernels to handwritten character recognition, on-line face recognition, and object detection from laser range finder sensor data. Parametric kernels achieve recognition rates competitive to state-of-the-art approaches in these tasks. / text

Tez (Yüksek Lisans) - Süleyman Demirel Üniversitesi, Fen Bilimleri Enstitüsü, Elektronik ve Haberleşme Mühendisliği Anabilim Dalı, 2007. / Kaynakça var.

Thesis (MScEng (Electrical and Electronic Engineering))--University of Stellenbosch, 2011. / ENGLISH ABSTRACT: Accurate water usage information is very important for municipalities in order to provide accurate billing information for high volume water users. Meter reading are currently obtained by sending a person out to every meter to obtain a manual reading. This is very costly with regards to time and money, and it is also very error prone. In order to improve on this system, an image based telemetry system was developed that can be retrofitted on currently installed bulk water meters. Images of the meter dials are captured and transmitted to a central server where they are further processed and enhanced. Character recognition is performed on the enhanced images in order to extract meter readings. Through tests it was found that characters can be recognised to 100% accuracy for cases which the character recognition software has been trained, and 70% accuracy for cases which is was not trained. Thus, an overall recognition accuracy of 85% was achieved. These results can be improved upon in future work by statistically analysing results and utilizing the inherent heuristic information from the meter dials. Overall the feasibility of the approach was demonstrated and a way forward was indicated. / AFRIKAANSE OPSOMMING: Dit is belangrik vir munisipaliteite om akkurate water verbruikingssyfers te hê sodat hulle akkurate rekeninge aan hoë volume water gebruikers kan stuur. Tans besoek ’n persoon fisies elke meter om meterlesings te verkry. Dit is egter baie oneffektief ten opsigte van tyd en geld. Die metode is ook baie geneig tot foute. Ten einde te verbeter op hierdie stelsel was ’n beeld gebaseerde telemetrie stelsel ontwerp wat geïnstalleer word op huidig geïnstalleerde hoë volume water meters. Beelde van die meters word na ’n sentrale bediener gestuur waar dit verwerk word en die beeld kwaliteit verbeter word. Karakter herkenning sagteware word gebruik om die meter lesings te verkry vanuit die verbeterde beelde. Deur middel van toetse is gevind dat karakters herken kan word tot op 100% graad van akkuraatheid in gevalle waar die karakter herkenning sagteware opgelei is, en 70% akkuraatheid vir gevalle waarvoor dit nie opgelei was nie. Dus was ’n algehele herkennings akkuraatheid van 85% behaal. Hierdie resultate kan verbeter word in die toekoms deur die resultate statisties te analiseer en die inherente heuristieke inligting van die meter syfers te benutting. Ten slotte, in die tesis was die haalbaarheid van die benadering gedemonstreer en ’n weg vorentoe vir toekomstige werk aangedui.