Scene Uyghur Character Detection System Based On Improved DBNet

After the invention of the written word,the written word serves as the carrier of recording and transmitting the information of human thoughts and feelings,and the written word can carry rich and accurate high-level semantic information.In natural scenes,the texture features and edge information of text,as a visual cue,are of great help to people's understanding of its semantic information,so how to better and faster access to the text information in the image has become very meaningful.Text Detection and recognition is the process of judging the position of a text and recognizing the meaning of a character by a certain algorithm.However,text detection and recognition in natural scenes is a great challenge due to the complex background,fuzzy fonts,occlusion and different language formats.Although text detection and recognition technology has made some achievements in recent years,and most of the research is conducted in English or Chinese,but for the Uyghur alphabets detection and recognition work less.The work of this paper includes the following four parts:(1)As a member of the OCR team at the key multilingual information technology laboratory in Xinjiang,the author was concerned with the lack of open and annotated Uighur data sets of natural scenes in the academic world,together with the team members in Xinjiang region to take photos and other means of real scene data collection,and carried out the follow-up image processing and tagging work,and thus a complete Uyghur alphabets image data set was established.The data set can be used for Uyghur character detection and recognition training,for which it not only marks text position but also marks text characters.(2)In view of the problems such as uneven illumination of text,irregular position change,complex background,small text and bending of text,which are difficult to detect in the Uyghur alphabets images of natural scenes at present,according to the structural characteristics of Uyghur alphabets,an improved DBNet scene detection algorithm was proposed,which could extract Uyghur characters effectively and detect small text,tilted text,curved text and so on.(3)For Representative Chinese and English detection networks in natural scenes,in this paper,CTPN network,R2 CNN network and DBNet network were selected to carry out a series of contrast experiments on the Uyghur alphabets data set presented in this paper,at the same time,based on the improved DBNet,the detection performance and effect were verified by the ablation experiment.The ablation experiment and the contrast experiment show that the improved DBNet takes into account the influence of small text and channel attention,the detection accuracy of the network was improved effectively.The accuracy and F value of the improved DBNet were improved by 3.16%and 7.43%,respectively.The detection accuracy based on the improved DBNet was76.72% and the F value was 67%.(4)According to the demand of Uyghur alphabets detection in natural scene,this paper designed and realized Uyghur alphabets detection system in natural scene image based on improved DBNet network.