Research On The Detection Of Altered Numerals In Handwriting Based On Image Classification

In the case of economic crimes,there is a lot of demand for altered numerals examination in handwriting.The identification results can not only restore the facts of the case,but also provide evidence for court trials.At present,the commonly used examination techniques for altered documents include microscopic examination,multi-band light source examination,infrared reflection photography examination,etc.However,due to the influence of equipment cost,professional skills and identification efficiency,the inspection is difficult and the degree of automation is low.In order to solve the above problems,this paper proposes an automatic detection method based on image classification for the altered numerals in handwriting.Two kinds of image classification technologies are used respectively: feature extraction + classifier design,and convolutional neural network to explore automatic detection methods suitable for altered numerals.The experimental results can provide reference for improving the intelligent level of document examination.In this paper,50 kinds of brand gel pens were used to write experimental samples on single background carrier(A4 paper)and complex background carrier(single receipt),respectively.Referring to several typical problems of modified numerals in judicial document expertise,the modified handwriting samples and normal handwriting samples of "3","4","6","7","8","9" were prepared.Used macro photography to collect images of samples,and preprocessed the collected original images in turn to obtain "WB-dataset(white background dataset)" and "CBdataset(complex background dataset)",with a total of 14,473 sample image sets.The weighted fusion feature algorithm and convolutional neural network were used for research respectively,and then aimed at the problems in the design of classical AlexNet network structure to build the network,and the network model FNNet(Forgery Numeral Net)suitable for altered numerals in handwriting detection was proposed and applied to case identification.The experimental results show that:(1)The average accuracy rate of the weighted fusion feature algorithm on two datasets reached 87.64% and 90.46%,respectively,indicating that the research on altered numerals detection based on image classification was feasible.(2)On the basis of AlexNet,the training parameters were adjusted by constantly training the network model,and the average accuracy rate of two datasets reached 95.35% and 94.61%,respectively,which was significantly better than that of the fusion feature algorithm.It proved that the convolutional neural network had great advantages in the detection of altered handwriting.(3)In view of the experimental results and existing problems of AlexNet,the improvement direction was put forward,and the network hierarchical structure parameters were selected according to the characteristics of the modified numerals,and FNNet network model was proposed.In the end,the average accuracy rate on the two altered numerals datasets was further improved,reaching 98.02% and 98.36% respectively,and the model parameters were reduced to 1/5 of AlexNet.The overall experimental results of the two datasets were also 3.21 percentage points higher than that of the AlexNet.(4)In order to verify the reliability of the research results in this paper,FNNet network model was used to test the altered numeral cases.The identification results were consistent with the facts,and automatic and rapid classification was realized,which further proved the practical application of this method.This paper breaks the traditional method of altered document examination,innovatively uses image classification technology to carry out experimental research,and finally proposes the FNNet network model for altered numerals detection,which provides a new automatic detection for altered document examination.