No abstract

No abstract

Offline handwriting recognition is a well-known challenging task in the optical character recognition field due to the difficulty caused by various unconstrained handwriting styles and limited training data. In order to learn invariant feature representations for handwriting, we propose a novel method to incorporate pixel-level rectification into a CNN-and RNNbased recognizer. We also propose an adjacent output mixup method for RNN layer's training to improve the generalization ability of the recognizer, i.e., the previous output of an RNN layer is added to the current output with random weights. We additionally adopt a series of techniques including pre-training, data augmentation and language model to significantly expand the training data scale, and further analyze their contributions to the improvement in the model performance. The proposed method performs well on four public offline handwriting benchmarks, including the IAM, Rimes, IFN/ENIT and CASIA-HWDB datasets.

Logical structure extraction of book documents is significant in electronic document database automatic construction.The tables of contents in a book play an important role in representing the overall logical structure and reference information of the book documents. In this paper, a new method is proposed to extract the hierarchical logical structure of book documents, in addition to the reference information, by combining spatial and semantic information of the tables of contents in a book. Experimental results obtained from testing on various book documents demonstrate the effectiveness and robustness of the proposed approach.

The research on offline handwritten Arabic character recognition has received more and more attention in recent years, because of the increasing needs of Arabic document digitization. The variation in Arabic handwriting brings great difficulty in character segmentation and recognition, eg., the subparts (diacritics) of the Arabic character may shift away from the main part. In this paper, a new probabilistic segmentation model is proposed. First, a contourbased over-segmentation method is conducted, cutting the word image into graphemes. The graphemes are sorted into 3 queues, which are character main parts, sub-parts (diacritics) above or below main parts respectively. The confidence for each character is calculated by the probabilistic model, taking into account both of the recognizer output and the geometric confidence besides with logical constraint. Then, the global optimization is conducted to find optimal cutting path, taking weighted average of character confidences as objective function. Experiments on handwritten Arabic documents with various writing styles show the proposed method is effective.