DEFESA DE TESE DE DOUTORADO Nº 29

Aluno: Arthur Flor de Sousa Neto

Título: "Overcoming Handwritten Data Scarcity: Synthesis and Recognition Using Residual Gated Convolution"

Orientador: Byron Leite Dantas Bezerra

Coorientador: Alejandro Héctor Toselli (NEU)

Examinador Externo: Thierry Paquet (URN)

Examinadora Externa: Nina Sumiko Tomita Hirata (USP)

Examinador Externo: Cleber Zanchettin (UFPE)

Examinador Interno: Bruno José Torres Fernandes

Data-hora: 15 de Abril de 2025 às 9h

Local: Formato Remoto - Google Meet.

---

Resumo:

         "Offline Handwritten Text Recognition (HTR) systems involve the automatic process of recognizing and transcribing handwritten text from scanned images into digital formats. The field has gained importance due to the increasing need for document digitization and the automation of data entry across various industrial sectors. However, achieving satisfactory recognition performance requires large and varied datasets for training optical models. The process of collecting and labeling such datasets is often time-consuming and impractical in many scenarios. To address this challenge, data augmentation is commonly applied; yet traditional augmentation methods may lead to model overfitting and performance degradation when data are scarce. Therefore, this work proposes integrating Conditional Generative Adversarial Networks (CGANs) for data synthesis into the optical model training to enhance handwriting recognition performance in data-scarce scenarios. To validate our proposal, we conducted a study that included: (i) a systematic literature review to identify gaps and trends in data augmentation for HTR; (ii) an exploration to establish an optimal configuration for traditional data augmentation; and (iii) extensive experiments using seven datasets. In addition, these datasets were partitioned into training subsets to simulate different data-scarce scenarios. The results indicate that, on average, data synthesis achieved a reduction of 40.1% in CER and 30.3% in WER, followed by transfer learning with reductions of 28.1% in CER and 21.4% in WER. In comparison, traditional data augmentation provided lower improvements, with average reductions of 19.3% in CER and 16.3% in WER. These findings demonstrate that both data synthesis and transfer learning enhance the performance of offline HTR systems, particularly in scenarios with limited training data. "