Feature Embedding Based Text Line Segmentation

In the process of social development,characters play an important role as information recording tools.With the development in the era of information,more requirements are put forward for the management and utilization of information.A large amount of handwritten document data is recorded in the form of document images.However simply saving document as photos is a very inefficient management mode.To facilitate the management of document information,it is necessary to establish a document image transcription,index and retrieval system,in which text line segmentation is still one of the most important preliminary tasks.In the task of text line segmentation,the text features are often variant.For instance,text lines significantly differ in text fonts,writing styles and font sizes.Moreover,the text lines are densely distributed.In some historical documents,there are interference of illustration,stains and other factors at the same time.With the development of deep learning method,it has made great progress in most fields of image processing.With the help of the powerful fitting ability of neural network,this thesis can make the model produce feature embedding with specific distribution through special design strategy.In this thesis,we propose a text line segmentation algorithm based on feature embedding.The main tasks of this thesis are as follows:First,under the basic feature embedding strategy,this thesis designs two text line segmentation algorithms based on different feature embedding strategies.In the high-dimensional feature embedding strategy,with the help of metric learning,this thesis realizes the task of text line segmentation by mapping the image to the high-dimensional space which can distinguish the adjacent text lines.In the low dimensional feature embedding strategy,this thesis uses the flow field to associate each position of the text area with its related text center line,to achieve the purpose of text line segmentation.Secondly,for two different feature embedding strategies of text line segmentation algorithm,the corresponding algorithm improvement strategy is proposed,and the effectiveness of the improved strategy is proved by ablation experiments.In the text line segmentation algorithm based on the high-dimensional feature embedding strategy,this paper designs a multi-center metric learning strategy to improve the high-dimensional feature embedding learned by the model,to improve the performance of the algorithm.In the text line segmentation algorithm based on the low dimensional feature embedding strategy,this thesis designs a self-tuning process to improve the quality of the flow field predicted by the model,and then improve the performance of the whole algorithm.Finally,with the help of comparative experiments,this paper compares the performance differences between the existing text line segmentation algorithm and the algorithm proposed in this paper on two open datasets.The experimental results show that in the text line segmentation problem,the performance of the text line segmentation algorithm based on the low-dimensional feature embedding strategy is significantly better than that based on the high-dimensional feature embedding strategy.Even compared with the existing text line segmentation algorithm,the proposed algorithm also achieved competitive performance improvement.