Research On Border Stamp Detection And Recognition Based On Deep Learning

Border stamps are stamps of state organs,and the inspection of border stamps is an important part of the border inspection procedure.It is one of the important procedures of identity verification and travel check for entry-exit personnel and plays an important role in maintaining border security and stability.In view of the increasing amount of data of border stamps,how to detect and identify border stamps accurately in real time is the main problems of border inspection.The existing manual verification method has low cost efficiency and some limitations,while the deep learning model has the advantages of fast,efficient and low cost.Therefore,this paper mainly studies a border stamps detection and recognition model based on deep learning.So that it can effectively detect the stamps in the passport in the actual scene,and output the country category of the stamps.Firstly,the stamp detection algorithm based on FECD object detection network is studied.On the one hand,the confidence loss function of FECD network is selected,the deficiency of cross entropy loss is analyzed,the focus loss is selected as the confidence loss of FECD network,which improves the model’s punishment for difficult samples and solves the problem of unbalanced samples between classes.On the other hand,the bounding box regression loss of FECD network is designed,and the ECDIo U bounding box regression loss is proposed.The overlap area between the prediction box and the ground truth box,the distance between the center points,the difference between the width and height and the distance between the four vertices were taken into full consideration.The prediction box of FECD network can accurately and efficiently approach to the ground truth box and improve the detection performance of the model.Stamp detection data set is constructed and experiments are carried out to verify the validity of FECD network design.After using the focus confidence loss function in the FECD network,the AP of the model is improved by 1.46% and the false detection rate on the test set is reduced by1.25%.On this basis,the AP of the model increases by 0.19% and the false detection rate decreases by 0.14% with the addition of ECDIo U bounding box regression loss.Then,the stamp recognition algorithm is studied based on AAR fine-grained classification network.Aiming at the problem that fine-grained feature extraction is not sufficient due to anchor generation mechanism,the aspect ratio of anchor generation mechanism is designed.k-means clustering algorithm is used to obtain the aspect ratio information of border stamp samples at different feature scales.Then it is used on the setting of aspect ratio of AAR network anchor generation mechanism.Therefore,the network can recognize the characters,patterns,outlines and other fine-grained features of stamps more effectively,and improve the fine-grained recognition effect of the model.The stamp recognition data set is constructed and experiments are carried out to verify the validity of the anchor generation mechanism design.The accuracy of AAR network model is improved by 0.26% after using the stamp aspect ratio obtained by clustering as the anchor aspect ratio.Finally,a stamp detection and recognition model is established to connect FECD object detection network with AAR fine-grained classification network.A single stamp sample output from the FECD network is taken as input to the AAR network.The location detection and country recognition of border stamps are successively implemented.The comprehensive accuracy of the final model is 97%.