Research On Text Detection And Recognition Algorithms For Tangut Ancient Books Based On Deep Learning

The Tangut ancient books are important documentary source for the study of the Tangut dynasty and have unique historical significance and cultural value.The intelligent text detection and recognition technology can digitalize images of Tangut ancient books,accurately detected location of Tangut ancient books in real time,and precisely translated Tangut ancient books.It will improve the research efficiency of Tangut scholars and will also effectively protect Tangut ancient books.Many Tangut ancient books have problems such as damage and yellowing due to historical changes,and Tangut characters have the characteristics of high similarity and complex strokes,resulting in a low recognition rate of Tangut ancient books.In this paper,we address the above problems and use deep learning algorithms to study text region detection and text recognition in Tangut ancient books.The main research content of this paper is as follows:(1)Collecting and annotating images of Tangut ancient books.Firstly we collect images of Tangut ancient books and pre-process them,and then annotate them and use data enhancement techniques to expand the dataset,and finally conduct statistical analysis of image size,aspect ratio,color bias distribution and sharpness distribution on the annotated text dataset of Tangut ancient books.(2)Improving the text detection model.As the traditional two-stage detection algorithm FasterRCNN faces the problem of detail loss when extracting features for small targets,the backbone network VGG16 is replaced with ResNet50,and the detection accuracy of the Tangut antique text is improved by optimizing the information processing method of the feature layer.However,the two-stage Faster-RCNN algorithm cannot meet the demand of real-time detection in terms of detection speed,so this paper focuses on the research and improvement of the one-stage target detection algorithm YOLOX.The CBAM attention mechanism is added after the convolutional layer in the CSPDarknet of YOLOX and the loss function is improved to finally achieve real-time and accurate detection of Tangut ancient texts.(3)Improving the text recognition model.The classification recognition model EfficientNet is improved to address the problem of low recognition rate of Tangut texts.Firstly,the multiple MBConv modules in the original network are simplified to reduce the computational effort.Then the ECA attention module is introduced to replace the SE attention module in the original network,which solves the dimensionality reduction problem caused by capturing non-linear cross-channel interactions,and also enables better capture of cross-channel interaction information,thus improving the accuracy rate of Tangut text recognition.