This paper presents an attention-based convolutional sequence to sequence (ACseq2seq) model for recognizing an input image of multiple text lines from Japanese historical documents without explicit segmentation of lines. The recognition system has three main parts: a feature extractor using Convolutional Neural Network (CNN) to extract a feature sequence from an input image; an encoder employing bidirectional Long Short-Term Memory (BLSTM) to encode the feature sequence; and a decoder using a unidirectional LSTM with the attention mechanism to generate the final target text based on the attended pertinent features. We also introduce a residual LSTM network between the attention vector and softmax layer in the decoder. The system can be trained end-to-end by a standard cross-entropy loss function. In the experiment, we evaluate the performance of the ACseq2seq model on the anomalously deformed Kana datasets in the PRMU contest. The results of the experiments show that our proposed model achieves higher recognition accuracy than the state-of-the-art recognition methods on the anomalously deformed Kana datasets.

Despite of recent breakthroughs in the accuracy of single character recognition using the deeper convolution neural networks, one of the remaining problems is that OCRs almost fail to recognize character patterns when they are severely degraded, especially those of the historical documents. Another problem to recognize characters in historical documents is the lack of sufficient training patterns because of the heavy cost for annotation. This paper proposes a character attention generative adversarial network named CAGAN for restoring heavily degraded character patterns in historical documents so that OCRs improve their accuracy and even help archeologists to decode them. The network is based on the U-Net like architecture [1] with skip connections, and it is trained by the proposed loss function including the common adversarial loss (global loss) and the hierarchical character attentive loss (local loss). We made an experiment on 118 categories of most common Japanese Kanji characters, collected from severely damaged historical documents called Heijokyo mokkan written during the Nara period in Japan. The experiment shows that our method restores the shapes of characters and improves the recognition rate significantly, which is helpful for archeologists to decode damaged character patterns.

This paper presents a semi-incremental recognition method for online Japanese handwritten text recognition, which is used for busy recognition interface (recognition while writing) and lazy recognition interface (recognition after writing) without large waiting time. We employ local processing strategy and focus on a recent sequence of strokes defined as "scope". For the latest scope, we build and update a segmentation and recognition candidate lattice and advance the best-path search incrementally. We utilize the result of the best-path search in the previous scope to exclude unnecessary segmentation candidates. This reduces the number of candidate character recognition with the result of reduced processing time. We also reuse the segmentation and recognition candidate lattice in the previous scope for the latest scope. Moreover, triggering recognition processes every few strokes save CPU time. Experiment made on TUAT-Kondate database shows the effectiveness of the proposed method not only in reduced processing time and waiting time, but also in recognition accuracy.