| Domestic ID cards are important identity certificates for Chinese citizens to conduct tourism,commerce and other activities within the country.With the increasingly frequent exchanges between the Mainland,Hong Kong,Macao and Taiwan,using the multi-spectral images to classify and authenticate ID cards efficiently and quickly is urgently needed.At present,there are few researches on this topic,and the classification and authentication algorithm for ID cards is faced with much difficulty,such as a wide variety of ID cards and complex backgrounds of ID card images.In view of the above problems,a classification and authentication algorithm for domestic ID cards based on multi-spectral images is proposed.(1)Aiming at the problems of a wide variety of ID cards,a classification algorithm for ID cards based on feature area detection and feature extraction is proposed.Firstly,a machine-readable zone detection algorithm is proposed to identify the face and orientation of ID cards.For one-line machine-readable code ID cards,a classification algorithm based on word area detection and gray-level co-occurrence matrix features is proposed.For three-line machine-readable code ID cards,a classification algorithm based on face area detection and improved Haar features is proposed.The algorithm decomposes the complex combination classification problem of ID card face,orientation,and type through a step-by-step refinement method,eliminates the background interference,and improves the accuracy of classification.(2)To address the problem of anti-counterfeiting areas of multi scales and rough edges in the segmentation of anti-counterfeiting areas,an anti-counterfeiting area segmentation algorithm for ID cards based on a dual-branch network with parameter sharing is proposed,which uses spatial paths and semantic paths to extract detailed information and semantic information respectively.A multi-scale module is proposed to obtain multi-scale features.The features are merged through the mutual guidance of the information of the two paths,which reduces the interference of background noise,improves the MIo U of multi-scale anti-counterfeiting areas.(3)Facing the problem of fake ID cards missing,an improved Con Sin Gan network is proposed.By increasing the number of convolutional layers at each stage and reducing the number of channels,it enhances the feature expression ability while avoiding the problem of overfitting and solving the problem of fake ID cards generation in a scenario where the background pix is small and the proportion of background is large.Aiming at the problem of different degrees of difficulty in authentication for ID cards,a progressive authentication algorithm based on pyramid structure is proposed.SSIM is used to quickly filter the easily distinguishable fake ID cards.The ID card authentication network is proposed to perform fine-grained authentication for fake ID cards with high similarity,which increases the intra-class similarity of real ID cards and reduces the similarity between real and fake ID cards,thereby reducing the MDR and UDR of the authentication algorithm for ID cards.The anti-counterfeiting area segmentation data set and authentication data set for ID cards are established based on the actual collected ID card images.The algorithm is implemented and tested to verify the feasibility and effectiveness of the classification and authentication algorithm for ID cards.Specifically,the accuracy rate of the face and orientation classification algorithm proposed is 98.75%,the accuracy rate of classification algorithm for one-line machine-readable code ID cards reaches 99.6%,and the accuracy rate of classification algorithm for three-line machine-readable code ID cards is 95%.The MIo U of the anti-counterfeiting area segmentation algorithm is90.33%,and the MDR and UDR of the authentication algorithm are 4.25% and 2.5%respectively. |
