The Banknote Detection Of Windowed Security Thread In Abnormal Situation Based On Deep Learning

The latest 2015 version of CNY firstly uses the anti-forgery technology of windowed security thread,and the special optically variable magnetic ink gives it obvious visual security feature,and the optically variable color has flowing effects viewed by the naked eye from different angles.It poses a great challenge to the existing CNY quality detection system in banknote industry.The unevenness,creasing of the windowed security thread and the unstable transmission will lead to abnormal phenomenon of reflection,which makes the quality detection on the windowed security thread not working.To solve this issue,the banknote detection method of windowed security thread in abnormal situation based on deep learning is proposed.The main research is as follow:(1)The window recognition of CNY windowed security thread.The experiment was carried out on the Tensor Flow Object Detection API,which adopted SSD-Mobile Net V1 as object detection model to realize the function of window localization.Firstly,the preparatory work such as image cropping,labeling was done to the collecting CNY images,and then I configured the training parameters and trained the recognition model of CNY with API.The recognition accuracy reached to 99.9%,which fully demonstrated the validity of the model.The window recognition model was applied to crop the window images from the windowed security thread of CNY,and prepared data for the following quality detection.(2)The hole detection on window images.Firstly,the classification model based on Bag of Words presenting for traditional object detection algorithm was designed for the experiments,which adopted the SIFT feature vectors extracted from the window images to build the visual dictionary based on K-means clustering algorithm,in order to represent the feature of window images and then send to SVM classification for training.Through adjusting the parameter of clustering center K,the average accuracy of the hole detection on the window could reach to 96.4%.Then the classification model based on residual neutral network presenting for deep learning algorithm was proposed,which could effectively detect the hole phenomenon by data preprocessing,building and training Res Net34,lateral comparison between models,comparison of the value of learning rate and so on,so that the accuracy reached up to 100% at the same dataset.The deep residual network standed out for its excellent learning ability and presentation capability.(3)The design and implementation of the detection system for windowed security thread of CNY.According to actual demand of quality process in banknote industry,this paper designed and implemented the banknote detection system of windowed security thread in abnormal situation with graphic interface,which was equipped with the function of readily training a network to meet production requirement,evaluating and inspecting the performance of training model,and detecting any CNY on windowed security thread.