Handwritten text recognition is an open problem of great interest in the area of automatic document image analysis. The transcription of handwritten content present in digitized documents is significant in analyzing historical archives or digitizing information from handwritten documents, forms, and communications. In the last years, great advances have been made in this area due to applying deep learning techniques to its resolution. This Thesis addresses the offline continuous handwritten text recognition (HTR) problem, consisting of developing algorithms and models capable of transcribing the text present in an image without the need for the text to be segmented into characters. For this purpose, we have proposed a new recognition model based on integrating two types of deep learning architectures: convolutional neural networks (CNN) and sequence-to-sequence (seq2seq) models, respectively. The convolutional component of the model is oriented to identify relevant features present in characters, and the seq2seq component builds the transcription of the text by modeling the sequential nature of the text. For the design of this new model, an extensive analysis of the capabilities of different convolutional architectures in the simplified problem of isolated character recognition has been carried out in order to identify the most suitable ones to be integrated into the continuous model. Additionally, extensive experimentation of the proposed model for the continuous problem has been carried out to determine its robustness to changes in parameterization. The generalization capacity of the model has also been validated by evaluating it on three handwritten text databases using different languages: IAM in English, RIMES in French, and Osborne in Spanish, respectively. The new proposed model provides competitive results with those obtained with other well-established methodologies.

This paper proposes the use of a new neural network architecture that combines a deep convolutional neural network with an encoder-decoder, called sequence to sequence, to solve the problem of recognizing isolated handwritten words. The proposed architecture aims to identify the characters and contextualize them with their neighbors to recognize any given word. Our model proposes a novel way to extract relevant visual features from a word image. It combines the use of a horizontal sliding window, to extract image patches, and the application of the LeNet-5 convolutional architecture to identify the characters. Extracted features are modeled using a sequence-to-sequence architecture to encode the visual characteristics and then to decode the sequence of characters in the handwritten text image. We test the proposed model on two handwritten databases (IAM and RIMES) under several experiments to determine the optimal parameterization of the model. Competitive results above those presented in the current state-of-the-art, on handwriting models, are achieved. Without using any language model and with closed dictionary, we obtain a word error rate in the test set of 12.7% in IAM and 6.6% in RIMES. (C) 2018 Elsevier B.V. All rights reserved.

The ATRECSIDE research project is investigating applications of deep learning models to automatic handwritten recognition problems, such as non-constrained extraction of text from document images, handwritten text recognition, and summarisation and prediction of texts.

Nowadays, Deep Learning is one of the most popular techniques which is used in several fields like handwriting text recognition. This paper presents our propose for a handwritten digit sequences recognition system. Our system, based in two stage model, is composed by Convolutional Neural Networks and Recurrent Neural Networks. Moreover, it is trained using on-demand scheme to recognize numbers from digits of the MNIST dataset. We will see that, with these training samples is not necessary segment or normalize the input images. Average recognition results were on 88,6% of accuracy in numbers of variable-length, between 1 and 10 digits. This accuracy is independent on the number length. Moreover, in most of the wrongly predicted numbers there was only one digit error.

We present our current work on building a deep learning architecture for the offline handwritten character recognition problem. The proposed system is based on training a deep Convolutional Neural Network (CNN) to recognize handwritten characters, using a new synthetic character database derived from UNIPEN dataset. The presented approach is inspired in some successfully-used neural architectures for image classification, specially the VGG-CNN. Our system reads each word with the help of a sliding window in a similar way to how humans do. An innovative feature of our proposal is using a synthetic character database specifically built, in a optimized way, to identify the characters as component elements of the words. Experiments with this new training synthetic dataset produced recognition rates of 98.4% for uppercase and 96.3% for lowercase, respectively.

In this paper, a methodology is proposed in order to produce a set of seeds later used as a starting point to K-Means-type unsupervised classification algorithms for text mining. Our proposal involves using the eigenvectors obtained from principal component analysis to extract initial seeds, upon appropriate treatment for search of lightly overlapping clusters which are also clearly identified by keywords. This work is motivated by the interest of the authors in the problem of identification of topics and themes previously unknown in short texts. Therefore, in order to validate the goodness of this method, it was applied on a sample of labeled e-mails (NG20) representing a gold standard within the field of text mining. Specifically, some corpora referenced in the literature have been used, configured in accordance to a mix of topics contained in the sample. The proposed method improves on the results of other state-of-the-art methods to which it is compared.

This paper presents a solution to the KDD CUP 2007 task "How Many Ratings?". The combination of three different approaches is used to produce a final solution which improves the results obtained by each of these procedures by itself.

The advance of the Internet as a competitor with traditional media (radio, TV, newspapers and magazines) is attracting advertisements, but traditional analytical tools for media planning are not directly applicable. In this paper we describe our experience in the Spanish TV domain and its evolution into the Internet arena.

This paper describes one possible way to solve task "Who rated what?" of the KDD CUP 2007. The proposed solution is a history-based model that predicts whether a user will vote a given movie. Key points to our approach are (1) the estimation of the model baseline, (2) the definition of the explanatory variables and (3) the mathematical model form. Given the binary outcome of the problem, the estimation of the true baseline (ratio of 1's in the test data) is critical in order to correctly make predictions. In parallel, to improve the model predictive power, we have developed a careful construction of the input variables. These explanatory variables can be grouped as: user voting behaviour variables, the movie characteristics and user-movie interactions. Finally, the mathematical model form (linear logistic regression) has been chosen among various model form competitors.