基于卷积神经网络的竖排版繁体中文图像文本化研究

Textualization of Traditional Chinese Images Based on Convolutional Neural Network

竖排版繁体中文图像文本化问题可以看作是古籍图像中文字的定位和识别问题,但目前主流的OCR技术对古籍文献中竖排版繁体中文的识别精度不高。为了解决此问题,本文将深度学习应用于文字识别和定位中。首先基于SSD模型,运用目标检测算法从古籍文献图像中定位文字;然后构建了Inception-Resnet卷积神经网络进行文字识别。测试表明,在同样数据集的条件下,针对不同版式、大小和字体的古籍文献图像,与其他经典神经网络模型相比,本文模型的性能和综合适用性更好。

The problem of textualization of vertical layout traditional Chinese images can be regarded as a problem of positioning and recognition of characters in images of ancient books,but the current mainstream OCR technology does not have high recognition accuracy forit. To solve this problem,this paper applied deep learning to text recognition and localization. Firstly,based on the SSD model,the target detection algorithm was used to locate the text from the ancient book document images;then the Inception-Resnet convolutional neural network was constructed for text recognition. The test showed that,under the same data set,the performance and comprehensive applicability of the model in this paper were better than other classical neural network models for ancient book images of different layouts,sizes and fonts.