Design And Implementation Of Text Recognition Algorithm For Travel Documents Image Base On Deep Learning

Travel documents are important identification documents for citizens of countries outside their borders,and commonly used travel documents include passports,laissezpasser,identity cards,etc.With the increasing frequency of international traffic,efficient automated identification of textual information on the surface of travel documents has become an urgent need for border clearance.Although a large number of research results on text detection and recognition have emerged in recent years,the complexity of the travel document’s background text,its varying scale and the diversity of languages make it difficult for existing algorithms to meet practical needs.To address the above problems,this paper proposes a text detection and recognition algorithm for travel document images.To address the problems of multiple languages,different text scales,and complex shading backgrounds for travel documents,this paper proposes a text detection algorithm based on multiple attention feature pyramid network.Through the fusion of multi-scale feature maps,the model’s precision and recall for multi-scale text are improved;by using multiple attention modules to model the spatial and channel context for feature maps,the semantic compactness and consistency of text features in different languages can be improved,and the feature differentiation between text regions and background regions can be improved,thereby improving the precision and recall of the text detection model for texts in multiple languages in complex backgrounds.To address the problem of complex text line image backgrounds and character adhesion in travel documents,this paper provides sequence recognition of the text line image as a whole to avoid complex single-character segmentation operations.In order to improve the recognition rate of the model for characters of different scales in a text line,a convolution module based on an adaptive receptive field is proposed;the accuracy and generalization ability of the text recognition model is improved by improving the connectionist temporal classification,improving the interference of the character class imbalance problem on the model training.The feasibility and effectiveness of the text detection and recognition algorithm proposed in this paper is verified by testing and comparing experiments using real-world captured travel document images with synthetic images.Specifically,the F1 score of the text detection algorithm proposed in this paper is 0.884,and the accuracy of the proposed text recognition algorithm is 0.986,which is higher than other popular methods.