María José Castro-Bleda

Holistic cursive word recognition based on perceptual features

This work presents a holistic system for the off-line recognition of cursive words based on the extraction of perceptual features by means of convolution with templates of line segments. The method works directly on gray-level images to avoid losing information due to binarization. Features are represented spatially in order to preserve topological information which is well suited for recognition through convolutional neural networks. Moreover, our feature extraction method does not need to detect baselines to find ascenders and descenders. The system has been tested on small and medium vocabulary sizes demonstrating the feasibility of the proposed method.

Fast Evaluation of Connectionist Language Models

Connectionist language models offer many advantages over their statistical counterparts, but they also have some drawbacks like a much more expensive computational cost. This paper describes a novel method to overcome this problem. A set of normalization values associated to the most frequent n -gramsis pre-computed and the model is smoothed with lower n -gramconnectionist or statistical models. The proposed approach is favourably compared to standard connectionist language models and with statistical back-off language models.

F-measure as the error function to train neural networks

Imbalance datasets impose serious problems in machine learning. For many tasks characterized by imbalanced data, the F-Measure seems more appropiate than the Mean Square Error or other error measures. This paper studies the use of F-Measure as the training criterion for Neural Networks by integrating it in the Error-Backpropagation algorithm. This novel training criterion has been validated empirically on a real task for which F-Measure is typically applied to evaluate the quality. The task consists in cleaning and enhancing ancient document images which is performed, in this work, by means of neural filters.

Improving a DTW-Based Recognition Engine for On-line Handwritten Characters by Using MLPs

Our open source real-time recognition engine for on-line isolated handwritten characters is a 3-Nearest Neighbor classifier that uses approximate dynamic time warping comparisons with a set of prototypes filtered by two fast distance-based methods. This engine achieved excellent classification rates on two writer-independent tasks:UJIpenchars and Pendigits. We present the integration of multilayer perceptrons into our engine, an improvement that speeds up the recognition process by taking advantage of the independence of these networks’ classification times from training set sizes. We also present experimental results on our new publicly available UJIpenchars2 database and on Pendigits.

Insights on the Use of Convolutional Neural Networks for Document Image Binarization

Convolutional Neural Networks have systematically shown good performance in Computer Vision and in Handwritten Text Recognition tasks. This paper proposes the use of these models for document image binarization. The main idea is to classify each pixel of the image into foreground and background from a sliding window centered at the pixel to be classified. An experimental analysis on the effect of sensitive parameters and some working topologies are proposed using two different corpora, of very different properties: DIBCO and Santgall.

Cache neural network language models based on long-distance dependencies for a spoken dialog system

The integration of a cache memory into a connectionist language model is proposed in this paper. The model captures long term dependencies of both words and concepts and is particularly useful for Spoken Language Understanding tasks. Experiments conducted on a human-machine telephone dialog corpus are reported, and an increase in performance is observed when features of previous turns are taken into account for predicting the concepts expressed in a user turn. In terms of Concept Error Rate we obtained a statistically significant improvement of 3.2 points over our baseline (10% relative improvement) on the French Media corpus.

Unconstrained offline handwriting recognition using connectionist character N-grams

This work presents an unconstrained offline hand-written line recognition system based on hybrid HMM (Hidden Markov Model)/ANN (Artificial Neural Network) models. The particularity of the system lies in the use of an ensemble of connectionist/statistical character n-gram language models. These language models are trained with a text corpus at character level; therefore, no explicit lexicon is used during recognition. The recognizer is thus able to output words which do not belong to that corpus. The proposed system favorably behaves compared to using a standard character n-gram on the IAM database lines corpus and achieves error rates comparable to state-of-the-art lexicon-driven alternatives.

Complete System for Text Line Extraction Using Convolutional Neural Networks and Watershed Transform

We present a novel Convolutional Neural Network based method for the extraction of text lines, which consists of an initial Layout Analysis followed by the estimation of the Main Body Area (i.e., the text area between the baseline and the corpus line) for each text line. Finally, a region-based method using watershed transform is performed on the map of the Main Body Area for extracting the resulting lines. We have evaluated the new system on the IAM-HisDB, a publicly available dataset containing historical documents, outperforming existing learning-based text line extraction methods, which consider the problem as pixel labelling problem into text and non-text regions.

Automatic speech segmentation based on acoustical clustering

In this paper, we present an automatic speech segmentation system based on acoustical clustering plus dynamic time warping. Our system operates at three stages, the first one obtains a coarse segmentation as a starting point to the second one. The second stage fixes phoneme boundaries in an iterative process of progressive refinement. The third stage makes a finer adjustment by considering some acoustic parameters estimated at a higher subsampling rate around the boundary to be adjusted. No manually segmented utterances are used in any stage. The results presented here demonstrate a good learning capability of the system, which only uses the phonetic transcription of each utterance. Our approach obtains similar results than the ones reported by previous related works on TIMIT database.

Cursive word recognition based on interactive activation and early visual processing models.

We present an off-line cursive word recognition system based completely on neural networks: reading models and models of early visual processing. The first stage (normalization) preprocesses the input image in order to reduce letter position uncertainty; the second stage (feature extraction) is based on the feedforward model of orientation selectivity; the third stage (letter pre-recognition) is based on a convolutional neural network, and the last stage (word recognition) is based on the interactive activation model.