Feature design and lexicon reduction for efficient offline handwriting recognition.

摘要

This thesis establishes a pattern recognition framework for offline word recognition systems. It focuses on the image level features because they greatly influence the recognition performance. In particular, we consider two complementary aspects of prominent features impact: lexicon reduction and the actual recognition. The first aspect, lexicon reduction, consists in the design of a weak classifier which outputs a set of candidate word hypotheses given a word image. Its main purpose is to reduce the recognition computational time while maintaining (or even improving) the recognition rate. The second aspect is the actual recognition system itself. In fact, several features exist in the literature based on different fields of research, but no consensus exists concerning the most promising ones. The goal of the proposed framework is to improve our understanding of relevant features in order to build better recognition systems. For this purpose, we addressed two specific problems: 1) feature design for lexicon reduction (application to Arabic script), and 2) feature evaluation for cursive handwriting recognition (application to Latin and Arabic scripts).;Few methods exist for lexicon reduction in Arabic script, unlike Latin script. Existing methods use salient features of Arabic words such as the number of subwords and diacritics, but totally ignore the shape of the subwords. Therefore, our first goal is to perform lexicon reduction based on subwords shape. Our approach is based on shape indexing, where the shape of a query subword is compared to a labeled database of sample subwords. For efficient comparison with a low computational overhead, we proposed the weighted topological signature vector (W-TSV) framework, where the subword shape is modeled as a weighted directed acyclic graph (DAG) from which the W-TSV vector is extracted for efficient indexing. The main contributions of this work are to extend the existing TSV framework to weighted DAG and to propose a shape indexing approach for lexicon reduction. Good performance for lexicon reduction is achieved for Arabic subwords. Nevertheless, the performance remains modest for Arabic words.;Considering the results of our first work on Arabic lexicon reduction, we propose to build a new index for better performance at the word level. The subword shape and the number of subwords and diacritics are all important components of Arabic word shape. We therefore propose the Arabic word descriptor (AWD) which integrates all the aforementioned components. It is built in two steps. First, a structural descriptor (SD) is computed for each connected component (CC) of the word image. It describes the CC shape using the bag-of-words model, where each visual word represents a different local shape structure. Then, the AWD is formed by concatenating the SDs using an efficient heuristic, implicitly discriminating between subwords and diacritics. In the context of lexicon reduction, the AWD is used to index a reference database. The main contribution of this work is the design of the AWD, which integrates lowlevel cues (subword shape structure) and symbolic information (subword counts and diacritics) into a single descriptor. The proposed method has a low computational overhead, it is simple to implement and it provides state-of-the-art performance for lexicon reduction on two Arabic databases, namely the Ibn Sina database of subwords and the IFN/ENIT database of words.;The last part of this thesis focuses on features for word recognition. A large body of features exist in the literature, each of them being motivated by different fields, such as pattern recognition, computer vision or machine learning. Identifying the most promising approaches would improve the design of the next generation of features. Nevertheless, because they are based on different concepts, it is difficult to compare them on a theoretical ground and efficient empirical tools are needed. Therefore, the last objective of the thesis is to provide a method for feature evaluation that assesses the strength and complementarity of existing features. A combination scheme has been designed for this purpose, in which each feature is evaluated through a reference recognition system, based on recurrent neural networks. More precisely, each feature is represented by an agent, which is an instance of the recognition system trained with that feature. The decisions of all the agents are combined using a weighted vote. The weights are jointly optimized during a training phase in order to increase the weighted vote of the true word label. Therefore, they reflect the strength and complementarity of the agents and their features for the given task. Finally, they are converted into a numerical score assigned to each feature, which is easy to interpret under this combination model. Five state-of-the-art features have been tested, and our results provide interesting insight for future feature design. (Abstract shortened by UMI.).