Convolutional Recurrent Network For Offline Handwritten Text And Scene Text Recognition

As a common media of social communication,texts have always been playing an important role in social development.And text recognition is still a challenging problem that needs more investigation.In this thesis,we mainly studied the task of recognizing handwritten and natural scene texts,both of which have their own difficulties in finding proper solutions..For handwritten texts,the changeable writing style and their cursive nature have caused considerable challenges for recognizing them.For scene texts,challenges are mainly arisen by the diversity of scenes and the natural environments that may put effects on the texts.The main work and contributions of this thesis include:1.A thorough introduction of common methods for off-line handwritten and natural scene text recognition.Based on the analysis of previous researches,our basic framework was constructed.2.For offline handwritten text recognition,we made a series of improvements upon the basic framework.Specifically,a multi-directional recurrent network module is proposed to learn the contextual information on various directions.Also,we incorporated the shortcut connection mechanism to take care of the cases where our deep network may incur bad convergence.These shortcut connection can also produce multi-level features in the network.3.For scene text recognition task,this thesis incorporated the attention mechanism network into the recognition framework.Through the visualization of weighted coordinate mapping,we analyzed the limitation of the basic attention network,and proposed the multi-rows attention model.In order to improve the accuracy of text localization,we proposed a learning algorithm for local region to help the network further extract the fine particle characteristics.