| With the development of the information age and the large-scale popularization of intelligent mobile terminals,online text recognition of mobile terminals is an important research direction in the field of artificial intelligence fusion pattern recognition,which has broad application prospects in the fields of machine translation,text recognition of image and video,and other text recognition in differentiated scenarios.In view of the current needs of research on the text recognition of Tibetan image on mobile terminals,taking realizing the text recognition APP of Tibetan images as a goal,this paper constructs a large-scale dataset of Tibetan text image annotation,studies the preprocessing and recognition technology of Tibetan image text combined with mobile terminal and system background,and builds the Tibetan image text recognition system APP on the Android platform,providing certain reference for future researches on Tibetan text recognition.The main research contents are as follows:（1）Analyzed the characteristics of the characters in Tibetan print and proposed a method of equal-cut segmentation for printed Tibetan.The mentioned method effectively solves the problem of the difficulties in segmentation caused by the conglutination of Tibetan characters in printed Tibetan.By automatically obtaining the width of each normal character image and cutting apart the character image of abnormal width equally according to the normal character width,the segmentation accuracy of Tibetan text character adhesion is effectively improved.（2）Trained and constructed an annotation dataset of Tibetan Text Image Character（abbreviation: TTIC）.As the deep learning model requires a lot of annotation data and there is currently no such publicly available annotation dataset,an annotation dataset of 229680 Tibetan character image was constructed through the combination of the proposed segmentation algorithm and manual work,which includes 638 character categories.（3）Researched and proposed the Tibetan print recognition method on the basis of CBAM-Le Net-5 model.In order to further improve the effect of recognizing Tibetan print,the attention weights in the two dimensions of feature graph channel and space are calculated in the convolutional network.The CBAM-Le Net-5 model enhances the ability of convolutional neural network（Le Net-5）to extract features,which performs better than the usual Alex Net and Res Net networks.Through the test,the final average recognition rate reached 96.03%.（4）Designed and established a text recognition system of Tibetan image based on Android platform,which includes processing function such as image acquisition,preprocessing,image text recognition and result display.The Tibetan text images of different Himalaya fonts were tested in the simulator and various real machine models.As a result,the APP ran well on the mobile terminal,and response time of each function is within 3 seconds.To sum up,this study realized the recognition system of Tibetan image text based on Android platform from the aspects of preprocessing technology combined with mobile terminal and system background,Tibetan text segmentation technology,and CBAM-Le Net-5 model of Tibetan print recognition,assisting to improve the experience of daily use by Tibetan compatriots. |
