结合密集神经网络与长短时记忆模型的中文识别

文本图像识别是计算机视觉领域一项重要任务，而其中的中文识别因种类繁多、结构复杂以及类间相近等特点很具挑战性.为改善这一问题，使用文本行端到端的识别模型.首次提出利用密集卷积神经网络（DenseNet）提取文本图像底层特征，同时避免手工设计、统计图像特征的繁琐；将整行图像特征直接送入双向长短时记忆模型（BLSTM）进行局部相关性分析，减少字符定位分割这一步骤；最后采用时域连接模型（CTC）解码获得识别的文本信息.实验表明所提出的模型可以高效的进行图像文本行的识别，并对图像的多种形变具有较好的鲁棒性.

Chinese Recognition Based on Dense Convolutional Network and Bidirectional Long Short-Term Memory Model

Text recognition is an important task in computer vision. The recognition of Chinese texts is challenging because of its wide range, complicated structure, and similar classes. In order to improve this problem, an end-to-end recognition model of text is used. The proposed model uses Dense convolutional Network (DenseNet) to extract features of text images, avoiding artificial design and statistics features. Then, the features are sent to Bidirectional Long Short-Term Memory model (BLSTM) for correlation analysis of local data. This step avoids the character segmentation. Finally, the Connectionist Temporal Classifier (CTC) is used to decode the text information. Experiments show that the proposed model can effectively recognize text images, and has strong robustness to various deformed images.