Identificação de escritores usando CNN's com abordagem de dissimilaridade

• Ano de defesa: 2019
• Autor(a) principal: Helal, Lucas Georges
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Estadual de Maringá; Brasil; Programa de Pós-Graduação em Ciência da Computação; UEM; Maringa; Centro de Ciências de Tecnologia
• Palavras-chave em Português: Convolutional Neural Network (CNN); Inteligência artificial; Rede neural; CNPQ::CIENCIAS EXATAS E DA TERRA; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: http://repositorio.uem.br:8080/jspui/handle/1/5361

The writer identification using handwritten documents has become an important research topic in documents forensics analysis. That is because it can be used as an identifying characteristic of a person. There are several databases composed of handwritten documents available for research, using different languages and alphabets. To obtain the objectives of this work, the following databases were utilized: CVL and BFL both used for single-script, for documents written in the same alphabet and the LAMIS-MSHD database which is built for multi-script, that is, documents written with differents alphabets. Besides that, several techniques were applied in writer identification process. The objective of this work is to evaluate the performance of the artificial intelligence technique known as Convolutional Neural Network (CNN) in writer identification utilizing handwritten documents. For this, a CNN will be used for the classification of the writers, also for feature extraction that will be evaluated in the SVM classifier and dissimilarity procedures are to be applied. Initially, experiments were developed using the traditional pattern recognition approach, based on feature engineering (or handcrafted features). In these experiments, the texture generation is done from the original documents and later, the features were extracted with the texture descriptors LBP and LPQ. Furthermore, the impact of the classification in the SVM was evaluated with and without the use of the dissimilarity approach, obtaining through combination rules of classifiers, a consensual decision in relation to the final decision. After a series of experiments, the approach with feature dissimilarity obtained through CNN, presented superior results in relation to the literature