Research On Scene Mongolian Character Detection And Recognition Based On Deep Learning

As the most direct representation form of human high-level semantic information,text plays an indispensable and important role in image understanding.In recent years,researchers have conducted in-depth research on the detection and recognition methods of English and Chinese characters in scene images,and have achieved fruitful results,but the research on Mongolian character detection and recognition methods in complex scene images is still in its infancy.In this context,the main research of this article are as follows:1.This paper studies a scene Mongolian text detection algorithm based on deep learning and maximum stable extreme value region(MSER).At present,due to the lack of large-scale scene Mongolian text detection database resources,it is impossible to train a robust deep detection network model,but the MSER-based method does not require a large number of training samples and is robust to changes in perspective,character size,and illumination.Based on the MSER method,relatively few training samples are needed to train an efficient and stable convolutional neural network(CNN)classifier,which is used to determine whether the candidate connected area is a Mongolian text area.Experimental results show that the proposed method can well complete the task of Mongolian text detection in scene images.2.This paper studies a method for generating Mongolian text samples in virtual scenes.In the rapid development of Mongolian language text information,a lot of work has been carried out on image analysis and text recognition of printed Mongolian documents,but less research has been carried out on Mongolian text recognition in scenes.On the one hand,the research on Mongolian language and script was carried out late,on the other hand,the lack of a large number of training samples made it impossible to directly apply deep learning methods to this field.The content of this research generates a virtual scene Mongolian text recognition training and test data set by simulating the interference of lighting,occlusion,complex background,deformation and other links that may be encountered in real scenes,in order to evaluate and improve existing methods.3.This paper studies a Mongolian text recognition method based on Siamese network.Mongolian writing is made by writing the Mongolian letters closely from top to bottom,and the internal local deformation is small,but the Mongolian writing category is large and the training samples are insufficient.These characteristics are similar to the face verification problem.Therefore,this study draws on the Siamese network method commonly used in face verification,and conducts a comparative test for the Mongolian text recognition problem in the scene,and converts the text recognition problem into the calculation of similarity between samples.Experimental results show that the proposed method can effectively alleviate the problem of insufficient training samples and improve the recognition performance of scene Mongolian characters.