Divulgação - Defesa Nº 173
Aluno: Dayvid Welles De Castro Oliveira
Título: “Deep Multidimensional Recurrent Neural Networks for Offline Handwriting Text Line Recognition”.
Orientador: Prof. Byron Leite Dantas Bezerra
Co-Orientador: Prof. Mêuser Jorge Silva Valença
Data-hora: 31/julho/2018 (9:00h)
Local: Escola Politécnica de Pernambuco – SALA I-4
Resumo:
“Handwriting still is an essential communication and documentation tool. Thus, in this digital era, it is common to find handwritten content recorded as image data. The off-line handwriting recognition is the task of converting a handwritten text image into a digital format that is understood by the computer. Currently, one of the main challenges in the handwriting recognition area lies in identifying complete lines of handwritten text. The advances brought by deep learning and especially the recent developments in recurrent neural networks as optical models in handwriting recognition systems have led to significant progress. However, we believe that there’s still room for improvements and one clear target is the architectural structure of these models. The main goal of this dissertation is to investigate alternative optical modeling approaches that can contribute to the optimization of offline and unconstrained handwriting recognition systems. In particular, we studied new architectural representations for a recognition system based on the Multidimensional Long Short-Term Memory (MDLSTM) in the hybrid Artificial Neural Network-Hidden Markov Model scheme. The MDLSTM architecture was elaborated to enhance the recognition performance and decrease the recognition time. Accordingly, we present modifications regarding the recurrent and convolutional aspects based on a state-of-the-art MDLSTM model. Since the results reported in the literature for deeper MDLSTM architectures relies on optimizing the network width with a fixed depth, we investigate the trade-off between both these hyper-parameters to obtain an optimal topology. The system was evaluated with English and French handwritten text lines from the IAM and RIMES databases respectively. As the main contribution, we show that the new hierarchical approach is able to maintain a robust recognition performance and still present significant speedups compared to a state-of-the-art MDLSTM architecture. The full handwriting recognition system including a decoder with linguistic knowledge presents competitive results compared to previous research on the considered benchmarks.”
O Programa
