Research On Offline Writer Identification Based On Deep Convolutional Neural Networks

Writer identification, considered as an important biometric identification technology, been widely used nowadays. Although many researchers have made breakthroughs on this field, the challenge of offline text-independent writer identification still remains, and there are many problems worth our further study.Deep Convolutional neural networks has been booming in recent years, and has solved many tough problems in computer vision, but little research has been reported on offline Chinese writer identification employing Deep convolutional neural networks. This paper focuses on offline text-independent Chinese writer identification and applies convolutional neural networks to this topic, which produces prominent results. The major contributions of the thesis are as follows:Above all. It's the first time for use deep convolutional neural networks to extract local features used for offline Chinese writer identification, and study several kinds feature encoding methods, and a global descriptor is then formed by means of Fisher Vectors encoding. Experiments show features extracted form convolutional neural networks are more distinctive than traditional expert-designed features.Then, several network structures is studied and optimized, so as to achieve better writer identification results. And the training of convolutional neural networks is probed into, a character segmentation method based on sliding window is proposed, and the effect of handwriting image normalization and data augmentation are evaluated to eventually achieve identification accuracy. In order to choose the more distinctive features, principal component analysis is adopted to select the feature extracted from convolutional neural networks, and the features of different network layers are compared.Lastly, In order to evaluate the proposed method a lot of careful experiments have been done. First, the experiments are conducted on Chinese dataset CASIA-HWDB, and produces the state of art results. Then, our method is tested on dataset ICDAR 2013 which contains two foreign languages, and also yields competitive results. The results reveal that the proposed method not only has good performance on Chinese writer identification, but also has a certain advantage for cross-language handwriting identification, and has strong adaptability and robustness.