This study investigated if it was feasible to use machine learning tools on OCR extracted text data to classify receipts and extract specific data points. Two OCR tools were evaluated, the first was Azure Computer Vision API and the second was Google Drive REST Api, where Google Drive REST Api was the main OCR tool used in the project because of its impressive performance. The classification task mainly tried to predict which of five given categories the receipts belongs to, and also a more challenging task of predicting specific subcategories inside those five larger categories. The data points we where trying to extract was the date of purchase on the receipt and the total price of the transaction. The classification was mainly done with the help of scikit-learn, while the extraction of data points was achieved by a simple custom made N-gram model. The results were promising with about 94 % cross validation score for classifying receipts based on category with the help of a LinearSVC classifier. Our custom model was successful in 72 % of cases for the price data point while the results for extracting the date was less successful with an accuracy of 50 %, which we still consider very promising given the simplistic nature of the custom model.

A way to follow products through the chain of production is important in the process industry and it is often solved by marking them with serial numbers. In some cases permanent markings such as dot peen marking is required. To ensure profitability in the industry and reduce errors, these markings must be read automatically. Automatic reading of dot peen markings using a camera can be hard since there is low contrast between the background and the numbers, the background can be uneven and different illuminations can affect the visibility. In this work, two different systems are implemented and evaluated to assess the possibility of developing a robust system. One system uses image analysis to segment the numbers before classifying them. The other system uses the recent advances in deep learning for object detection. Both implementations are shown to work in near real-time on a cpu. The deep learning object detection approach was able to classify all numbers correct in a image 60% of the time, while the other approach only succeeded in 20% of the time.

Les recherches et les logiciels présentés dans cette étude s'adressent à une importante minorité au sein de notre société, à savoir la communauté des sourdes. De nombreuses recherches démontrent que les sourdes se heurtent à de grosses difficultés avec la langue vocale, ce qui explique pourquoi la plu- part d'entre eux préfère communiquer dans la langue des signes. Du point de vue des sciences de l'information, les LS constituent un groupe de minorités linguistiques peu représentées dans l'univers du numérique. Et, de fait, les sourds sont les sujets les plus touchés par la fracture numérique. Cette étude veut donc être une contribution pour tenter de resserrer cette fracture numérique qui pénalise les sourdes. Pour ce faire, nous nous sommes principalement concentrés sur l'informatisation de SignWriting, qui constitue l'un des systèmes les plus prometteurs pour écrire la LS. / The studies and the software presented in this work are addressed to a relevant minority of our society, namely deaf people. Many studies demonstrate that, for several reasons, deaf people experience significant difficulties in exploiting a Vocal Language (VL English, Chinese, etc.). In fact, many of them prefer to communicate using Sign Language (SL). As computer scientists, we observed that SLs are currently a set of underrepresented linguistic minorities in the digital world. As a matter of fact, deaf people are among those individuals which are mostly affected by the digital divide. This work is our contribution towards leveling the digital divide affecting deaf people. In particular, we focused on the computer handling of SignWriting, which is one of the most promising systems devised to write SLs.

Context. Custom solutions to optical character recognition problems are able to reach higher recognition rates then a generic solution by their ability to exploiting the limitations in the problem domain. Such solutions can be generated with genetic algorithms. This thesis evaluates two different chromosome encodings on an optical character recognition problem with a limited problem domain. Objectives. The main objective for this study is to compare two different chromosome encodings used in a genetic algorithm generating neural networks for an optical character recognition problem to evaluate both the impact on the evolution of the network as well as the networks produced. Methods. A systematic literature review was conducted to find genetic chromosome encodings previously used on similar problem. One well documented chromosome encoding was found. We implemented the found hromosome ncoding called binary, as well as a modified version called weighted binary, which intended to reduce the risk of bad mutations. Both chromosome encodings were evaluated on an optical character recognition problem with a limited problem domain. The experiment was run with two different population sizes, ten and fifty. A baseline for what to consider a good solution on the problem was acquired by implementing a template matching classifier on the same dataset. Template matching was chosen since it is used in existing solutions on the same problem. Results. Both encodings were able to reach good results compared to the baseline. The weighted binary encoding was able to reduce the problem with bad mutations which occurred in the binary encoding. However it also had a negative impact on the ability of finding the best networks. The weighted binary encoding was more prone to enbreeding with a small population than the binary encoding. The best network generated using the binary encoding had a 99.65% recognition rate while the best network generated by the weighted binary encoding had a 99.55% recognition rate. Conclusions. We conclude that it is possible to generate many good solutions for an optical character problem with a limited problem domain. Even though it is possible to reduce the risk of bad mutations in a genetic lgorithm generating neural networks used for optical character recognition by designing the chromosome encoding, it may be more harmful than not doing it.

Optical Character Recognition (OCR) is a well studied subject involving variousapplication areas. OCR results in various limited problem areas are promising,however building highly accurate OCR application is still problematic in practice.This thesis discusses the problem of recognizing and confirming Bingo lottery numbersfrom a real lottery field, and a prototype for Android phone is implementedand evaluated. An OCR library Tesseract and two Artificial Neural Network (ANN)approaches are compared in an experiment and discussed. The results show thattraining a neural network for each number gives slightly higher results than Tesseract.

D.Ing. (Electrical and Electronic ) / This thesis describes a neural network based system for the classification of handwritten digits as found on real-life mail pieces. The proposed neural network uses a modular architecture which lends itself to parallel implementation. This modular architecture is shown to produce adequate performance levels while significantly reducing the required training time. The aim of the system is not only to achieve a high recognition performance, but also to gain more insight into the functioning of the neural networks. This is achieved by using separate feature extraction and classification stages. The output of the feature extraction stage gives a good indication of the final performance level of the classifier, even before training. The need for an optimal feature set is expressed to elevate the performance levels even further.

D. Ing. / This thesis describes a real world application in the field of pattern recognition. License plate recognition and face recognition algorithms are combined to implement automated access control at the gates of RAU campus. One image of the license plate and three images of the driver’s face are enough to check if the person driving a particular car into campus is the same as the person driving this car out. The license plate recognition module is based on learning vector quantization and performs well enough to be used in a realistic environment. The face recognition module is based on the Bayes rule and while performing satisfactory, extensive research is still necessary before this system can be implemented in real life. The main reasons for failure of the system were identified as the variable lighting and insufficient landmarks for effective warping.

This thesis presents the design of a recognizer for unconstrained handprinted alphanumeric characters. The design is based on a thinning process that is capable of producing thinned images with well defined features that are considered essential for character image description and recognition. By choosing the topological points of the thinned ('line') character image as these desired features, the thinning process achieves not only a high degree of data reduction but also transforms a binary image into a discrete form of line drawing that can be represented by graphs. As a result powerful graphical analysis techniques can be applied to analyze and classify the image. The image classification is performed in two stages. Firstly, a technique for identifying the topological points in the thinned image is developed. These topological points represent the global features of the image and because of their invariance to elastic deformations, they are used for image preclassification. Preclassification results in a substantial reduction in the entropy of the input image. The subsequent process can concentrate only on the differentiation of images that are topologically equivalent. In the preclassifier simple logic operations localized to the immediate neighbourhood of each pixel are used. These operations are also highly independent and easy to implement using VLSI. A graphical technique for image extraction and representation called the chain coded digraph representation is introduced. The technique uses global features such as nodes and the Freeman's chain codes for digital curves as branches. The chain coded digraph contains all the information that is present in the thinned image. This avoids using the image feature extraction approach for image description and data reduction (a difficult process to optimize) without sacrificing speed or complexity. After preclassification, a second stage of the recognition process analyses the chain coded digraph using the concept of attributed relational graph (ARG). ARG representation of the image can be obtained readily through simple transformations or rewriting rules from the chain coded digraph. The ARG representation of an image describes the shape primitives in the image and their relationships. Final classification of the input image can be made by comparing its ARG with the ARGs of known characters. The final classification involves only the comparison of ARGs of a predetermined topology. This information is crucial to the design of a matching algorithm called the reference guided inexact matching procedure, designed for high speed matching of character image ARGs. This graph matching procedure is shown to be much faster than other conventional graph matching procedures. The designed recognizer is implemented in Pascal on the PDP11/23 and VAX 11/750 computer. Test using Munson's data shows a high recognition rate of 91.46%. However, the recognizer is designed with the aim of an eventual implementation using VLSI and also as a basic recognizer for further research in reading machines. Therefore its full potential is yet to be realized. Nevertheless, the experiments with Munson's data illustrates the effectiveness of the design approach and the advantages it offers as a basic system for future research. / Applied Science, Faculty of / Electrical and Computer Engineering, Department of / Graduate

This paper describes basic methods used for optical character recognition. It explains all procedures of recognition from adjustment of picture, processing, feature extracting to matching algorithms. It compares methods and algorithms for character recognition obtained graphically distorted or else modified image so-called „captcha“, used in present. Further it compares method based on invariant moments and neural network as final classifier and method based on correlation between normals and recognized characters.

This thesis aims to implement a segmentation-free strategy in the context of handwritten multi-digit string recognition. Three models namely VGG-16, CRNN and 4C are built to be evaluated and benchmarked, also research about the effect of the different training set on model performance is carried out.