Method for positioning and identifying electronic image archiving stamp of paper archive

Abstract

The invention relates to a method for positioning and identifying an electronic image archiving stamp of a paper archive. The method comprises the following steps: marking stamp training data: selecting a fixed part and an exhaustible part as characteristics of archiving stamp detection to construct a training set; configuring a model: adopting ResNet-101 as an improved SSD network model of the pre-feature extraction network; training a model: training the improved SSD network model by adopting the training set; and using the model: inputting a to-be-decrypted electronic document into the trained improved SSD network model, outputting a boundary region of a positive sample, and judging whether a security classification archiving seal exists or not according to the value of the positive sample. The detection precision of the archiving seal can be improved.

Description

technical field [0001] The invention relates to the technical field of seal detection, in particular to a method for locating and identifying paper archive electronic image archive stamps. Background technique [0002] After the paper archives exceed the time limit stipulated by the classification level, in line with the principle of whoever decides the secret, whoever decrypts it, they can be made public and reused after going through the declassification review process of the classification unit. When using the electronic images of these paper archives, they also need to be "decrypted" to avoid misunderstandings when they are used by the public. [0003] At present, the positioning and identification of electronic image archive seals of paper archives mainly include the following methods: manual identification and detection methods, computer pattern recognition methods, and machine learning methods. [0004] The manual identification detection method is mainly through man...

AI Technical Summary

Problems solved by technology

However, the SSD model also has some deficiencies. For example, the gradient explosion caused by the SSD model feature sampling backpropagation makes network training difficult; even if the training can continue to be completed at a small layer depth, the clarity will appear after a certain number of training iterations. Saturation, resulting in a decrease in accuracy will still have a very large impact on the output

Images