Classificação de pontos de segmentação de dígitos manuscritos

AUTOR(ES)

Eduardo Vellasques

DATA DE PUBLICAÇÃO

2006

RESUMO

This work presents a method to classify segmentation cuts for handwritten digits. The proposed method works as a filter, which is applied on segmentation-based recognition systems. In this strategy, the number of segmentation hypothesis created is usually high. These segmentation hypothesis are individually evaluated by a generalpurpose classifier. Filtering unnecessary cuts, not only reduces the computational effort (as the computational cost of the proposed filter is smaller than the computational cost of the generalpurpose classifier) but also increases the recognition rate of the general-purpose classi fier. Through the reduction in the number of segmentation hypothesis, many calls to the general-purpose classifier, which works with 132 features are migrated to a segmentation cut classifier (filter) which works with only 42 features. Moreover, in the proposed strategy, this is done with a significantly smaller number of calls to the filter (when compared to the number of calls to the general-purpose classifier). High discriminant features (MCA) were used. This type of feature relies on concavity analysis before and after the segmentation is done. This makes possible the detection of over-segmentation within a given connected component. Detecting oversegmentation was the way chosen to detect unnecessary cuts, as an they usually generate over-segmentation. Besides, the chosen feature set is invariant to the number and type of segmentation cuts, string length. This makes possible the use of the proposed method in every system which relies on segmentation-based recognition. The use of an SVM classifier (to detect over-segmentation) and of ROC analysis (to optimize the performance of the proposed filter) makes the proposed method adaptive and heuristics-free. The experimental results show us that the proposed method makes possible to reduce the computational cost and to increase the recognition performance, at the same time.

ASSUNTO(S)

processamento de imagens informática roconheciemnto de padrões ciencia da computacao

Documentos Relacionados

• HANDWRITTEN DIGITS RECOGNITION BY NEURAL NETWORKS
• Segmentação automatica e treinamento discriminativo aplicados a um sistema de reconhecimento de digitos conectados
• Segmentação automática e classificação em vídeos esportivos
• Classificação de pontos LIDAR para a geração do MDT
• Quantificação do colágeno dérmico equino por duas técnicas morfométricas: contagem de pontos e segmentação de cor