Text Detection Of Tangut Based On Deep Learning

The ancient books of Xixia recorded many political,economic,military,cultural and other events in the Xixia period.They are important documents for related researchers to interpret the Xixia civilization and have high historical value.Due to the long preservation time and improper preservation methods,the ancient texts have yellowing,missing,decayed and other phenomena,which have a certain impact on the detection and recognition of ancient texts.This paper takes Xixia ancient texts as the research object,compares traditional techniques,and uses deep learning algorithms under different frameworks to study the detection and positioning of text regions in ancient documents.The main work of the thesis is as follows:Firstly,collect a considerable number of text images of ancient books in Xixia,mark and locate the text area in the images of ancient books and establish a text data set of Xixia ancient books using the conventional data set expansion method.At the same time,statistics are made on factors such as the length and width of the marked text area and the area ratio,and further analysis of the difference between it and the conventional target and its own characteristics,to provide a certain direction for follow-up research.Secondly,on the basis of Xixia text as the detection target,the two-stage R-CNN series algorithm based on Anchor-Base is compared with the first-stage SSD and YOLO series algorithms,and it is analyzed that each algorithm detects the existence of this type of target.The problem,and the corresponding improvement plan is given as follows.Aiming at the problem that the SSD algorithm network is relatively shallow,the feature layer is insufficient,and the priori box generation method based on general target detection is unreasonable in the field of text target detection.Then combined with the statistical data of the length and width of the Xixia text,the algorithm is optimized the generation method of the priori box,using the hollow convolution instead of the pooling layer,combined with the idea of feature fusion and attention mechanism,effectively improves the algorithm’s detection accuracy of Xixia text targets.Considering the irrationality of the YOLOv3 algorithm in judging the positive and negative samples of the text target,combined with the length,width,location and other information of the text target,the CIOU algorithm is used as the standard for judging the positive and negative samples.For the unity and solidity of the algorithm feature layer fusion module,the FSKFPN feature fusion module is proposed and compared with the adaptive feature fusion module for experimental comparison and analysis.Considering the actual application scenarios of text detection,the text detection algorithm E-YOLO based on restricted environments is proposed,and the Efficient-Net series network is used as the backbone feature extraction network to reduce the amount of network calculations.In the case of a small loss of detection accuracy,a high detection speed is maintained.Then,analyze the common problems of each algorithm based on the idea of Anchor-Base—the algorithm prediction frame does not fit well with the target actual frame.The CenterNet algorithm of Anchor-Free idea is used to effectively solve the problem of poor fit between the prediction frame and the actual target,and the hybrid space domain module is used to strengthen the algorithm to extract effective feature information.At the same time,the structure of U-Net is used for reference to alleviate the feature of the network.The problem of semantic ambiguity improves the detection ability of the algorithm.Finally,according to the text target location information detected by various algorithms,the text column is segmented,and a single column of Xixia ancient text pictures is cut out,and a variety of image binarization methods are used for comparison to reduce the noise of the cut pictures.Through a series of operations such as detection,positioning and segmentation of ancient texts in Xixia,a solid foundation has been laid for the subsequent research on Xixia text recognition.