The Research Of Banknotes Discrimination Based On Infrared Characteristics

In the circulation of paper money in the market,it is inevitable that counterfeit money will be mixed with it,and the number of counterfeit currencies is small and unbalanced.This kind of small sample imbalanced classification learning is currently a key breakthrough in the field of artificial intelligence.It is widely used in deep learning for image recognition,and it is also prone to overfitting for this kind of banknote data classification.In view of this situation,it is necessary to study an intelligent predictive counterfeiting algorithm that solves small sample learning.The research of this paper is based on the theoretical research on the problem that the existing counterfeit currency samples are small and unbalanced.The design algorithm performs preprocessing operations of tilt correction,filtering and denoising and cutting of samples.A new infrared banknote verification model is proposed and optimized and improved based on the proposed model.Research has shown that the model solves the problem of imbalanced samples and small sample infrared banknote image recognition to a certain extent,and also provides a reference direction for this type of image recognition.The main innovations of this paper are as follows:(1)In-depth study of infrared image noise processing,a combination of wavelet transform and Gabor algorithm is proposed to denoise the image,the high-frequency detail sub-images of the image after wavelet transform are fused with Gabor filter transform,and then reconstructed by wavelet transform to obtain denoising Image.The method achieves denoising while retaining the detailed information of the banknotes,which is convenient for subsequent banknote recognition.(2)In-depth study of generative adversarial network theory,a new infrared banknote verification model based on semi-supervised auxiliary classification and generation confrontation network is proposed to complete banknote feature extraction and recognition.The model uses the generator and discriminator game to cascade the label and random noise to generate a picture with the characteristics of our sample,and use unlabeled data to further optimize the discriminator and expand our sample number.The discriminator formed after the feature extraction of banknotes can not only judge the authenticity of the image but also classify the banknotes.After adding the classification,the ability of the discriminator generator is also improved,and the ideal classification recognition is realized in the dynamic optimization.The expanded sample is used to further optimize the discriminator,so that a small number of label samples can use the game between the generator and the discriminator to improve the model feature extraction ability,optimize network parameters,balance the number of samples,continue to optimize classification,and complete banknote recognition.(3)Deeply study the basic principles of the attention mechanism and make improvements to the model proposed above.Attention is introduced into the discriminator of the generated adversarial network,and a weight prediction branch is added to the discriminator convolution channel,and it is expected to automatically learn the importance of channel features.Through global average pooling of the input feature map,and then a sigmoid function,the weight vector of the channel feature is obtained,and then the original feature map in the channel is multiplied and weighted to achieve the purpose of automatically focusing on the channel with a large amount of information.Based on the improved auxiliary classification depth generation adversarial network algorithm to realize banknote recognition,the accuracy rate has been improved under certain conditions,but there is no significant change in the accuracy rate at the initial training stage.