Research On The Influence Of Stamping Conditions On Automatic Recognition Of Stamp Impression

With the continuous development of computer technology,in the field of image recognition,automatic recognition of stamp impression has dawned.Among the many methods of automatic recognition,one based on neural network has greater advantages,which has been demonstrated that neural networks can be used to distinguish the stamp impression stamped by different seals through studies.However,in the field of forensic examination of document,no researchers have analyzed the influence of stamping conditions on automatic recognition of stamp impression from the actual needs of forensic identification of stamp impression,let alone the exploration to the feasibility and practicability of using convolutional neural network to recognize stamp impression.In this case,we were focusing on the study of the influences of stamping conditions on automatic recognition of stamp impression by using convolutional neural networks and the feasibility and practicability of using convolutional neural network to recognize stamp impression.The research was conducted mainly on following aspects:(1)Three seals with universal use were selected(rubber seal,copper seal and photosensitive seal).Under different stamping conditions,450000 stamp impression were manually stamped as experimental samples.Four representative convolutional neural network models(Lenet-5,Alexnet,VGG16,Resnet50)were selected for the pre-experiment.The experimental results show that:(1)the size of training samples is positively related to the recognition accuracy.(2)The optimal training samples of four network models are obtained.(3)Under ideal conditions,Resnet50 has a better recognition effect.(2)By changing the shades of stamp impression,the types of printing materials,the integrity of stamp impression in training and testing samples,to probe into the influence of stamping condition on the recognition of stamp impression by convolutional neural networks.The experimental results show that the stamping condition has a great influence on the automatic recognition of stamp impression.When the stamping conditions of training and testing samples are the same,the recognition accuracy could be close to or up to 100%.When the stamping conditions of training and testing samples are different,the recognition accuracy is significantly reduced.As the stamp impression with multiple stamping conditions are used as training sets to recognize the stamp impression with multiple stamping conditions,the accuracy could be close to or up to 100%.(3)The feasibility and practicability of using convolutional neural network to recognize stamp impression are discussed.The experimental results show that,by convolutional neural network,under the condition of enough training samples and sufficient stamping conditions,recognizing the true and false stamp impression has high accuracy rate.The method is feasible and can be used as a reference for forensic identification of stamp impression.However,when the number of training samples is small,the recognition accuracy is relatively low,even wrong recognition results appear.Due to the limited recognition accuracy and the long preparation time of preparing training samples,the practicability of the method is poor.