Certificate Image Detection And Text Recognition For Mobile Terminal

Text is the source of perceptual information closely related to social life.Compared with other content in the perceptual image,text contains more concise semantic information.With the development and innovation of 5G technology and deep vision network,the application environment and life of various emerging artificial intelligence are intertwined and integrated.The automatic image processing system of PC terminal is gradually transferring to intelligent terminal equipment,and the image data resources with text information are showing explosive growth.Retrieving target text information from massive image carriers has become a hot research direction,Automatic recognition and extraction of image text information also has a wide range of business needs and application prospects.In view of the certificate image as the evidence document commonly used in identity authentication,medical and health,business processing,census and other affairs,this paper studies a method of certificate image detection and text recognition suitable for mobile terminals.The main content includes two parts:(1)Methods of certificate image detection and correction.The detection and correction of the certificate image captured by the mobile terminal is often affected by the background environment,shooting angle,camera resolution,light interference and other factors.It usually has the characteristics of poor imaging quality,text distortion,variable angle tilt,unclear text and so on.These characteristics become the uncertain factors that affect the performance of the system,This paper presents a simple and robust method for certificate image detection and character recognition,which is suitable for mobile terminals.Through a series of image processing methods,such as foreground segmentation,contour extraction,fitting lines,seeking corners,affine transformation,the certificate is extracted and corrected.The test results show that this method has good robustness and the corrected image and character are not distorted or deformed,The accuracy of the subsequent recognition process is significantly improved.(2)Certificate image recognition without detection based on transformer.The existing text recognition models need to rely on two processes of detection and recognition,and need the help of complex network structure,a large number of text box labels and training sets to improve the recognition accuracy,which makes the network model with large amount of calculation,high complexity and easy to produce cumulative error.In order to solve the above technical problems,this paper improves the model architecture of transformer,adds the global context module and proposes a learnable two-dimensional position coding.By embedding the position coding of image feature graph,the feature expressions of different subspaces are directly connected to the sequence decoder.The model can be trained in parallel and converges rapidly,By inserting special symbols into text annotation,structured field data can be directly obtained and automatically filed.According to the experimental results of ID card,the recognition rate of the network model in gender,nationality,birth date and ID number Ordinance has reached over 98%.Compared with the original Transformer benchmark,the recognition rate of name and residential address has reached 86.1% and 98.8%,respectively,which has increased by 17.7% and 8.5% respectively.In this paper,a simple and robust method of certificate image detection and character recognition for mobile terminals is proposed.The certificate in the image is extracted and corrected by image segmentation and correction method,and then the corrected image is recognized.The test results show that the proposed method has excellent performance in the identification of certificate images captured by mobile terminals.