An Automatic Detection Method for Forged Fingerprints Based on Optical Coherence Tomography

Abstract

An automatic detection method for forged fingerprints based on optical coherence tomography, comprising the following steps: 1) collecting a sequence of fingerprint slice images through an Optical Coherence Tomography (OCT) device; 2) preprocessing the fingerprint slice images, including image denoising and image size adjustment; 3) reconstructing through a memory-enhanced denoising autoencoder network to obtain the reconstruction error of the latent space features; 4) judging whether the image blocks of the fingerprint slices are real fingerprints according to a threshold, so as to realize the automatic detection of forged fingerprints. The present invention effectively solves the problems such as the dependence of the supervised network model training method on data annotation and the poor robustness of the artificially designed anti-counterfeiting feature detection, and at the same time improves the accuracy and generalization ability of the automatic detection method for forged fingerprints.

Description

Technical Field

[0001] The present invention belongs to the field of biometric recognition security, and particularly relates to an automatic forged fingerprint detection method based on Optical Coherence Tomography (OCT), which is a kind of deep learning technology and belongs to the method of learning without labeled data. Background Art

[0002] Fingerprint is a reliable biometric recognition technology. Identity verification systems based on fingerprint recognition are widely used in scenarios with high security requirements such as mobile payment and border control. However, the recognition system still has the risk of being attacked by forged samples, such as 3D printed fingerprints and artificial fingerprints made of low-cost materials such as silicone and resin. There is an urgent need for robust fingerprint authenticity detection methods as the first line of defense to ensure system security. The research on fingerprint anti-counterfeiting technology can be classified into two categories: hardware and software.

[0003] The hardware-based method requires updating the fingerprint acquisition device and adding a new type of sensor to detect the authenticity of fingerprints by detecting live body features, such as blood flow, skin distortion, chemical composition, etc. Compared with other imaging technologies, OCT imaging has high resolution and can detect the additional covered cortex and epidermal fingerprints on the finger, as well as the micron-level internal complex structure of multiple layers of tissues, which can provide more reliable anti-counterfeiting performance. The software method extracts features from the fingerprint images obtained from the fingerprint sensor and designs different anti-counterfeiting methods for different fingerprint forgery technology attacks to distinguish live and forged fingers. However, the forms of fingerprint forgery are diverse, and it is difficult for supervised and human-intervened methods that rely on data labels to identify unseen types of forged fingerprints. Summary of the Invention

[0004] In order to overcome the data dependence of the supervised learning method and the problem that the recognition algorithm fails in the face of new forgery means, and further improve the generalization ability of the automatic forged fingerprint detection method, the present invention proposes a memory-enhanced denoising autoencoder network. After image preprocessing of the real fingerprint slice images collected by the OCT system, a local area is intercepted and input into the encoder network for reconstruction training. The trained network model can quickly and automatically detect forged fingerprints by obtaining the feature reconstruction error of the sample to be tested.

[0005] The technical solution adopted by the present invention to solve the above problems is:

[0006] An automatic forged fingerprint detection method based on optical coherence tomography, the method comprising the following steps:

[0007] 1) Collect a sequence of fingerprint slices through an Optical Coherence Tomography (OCT) device;

[0008] 2) Perform image preprocessing on the fingerprint slice image;

[0009] 3) Reconstruct through a memory-enhanced denoising autoencoder network to obtain the reconstruction error of the latent space features;

[0010] 4) Determine whether the image block of the fingerprint slice is a real fingerprint according to the threshold, so as to realize the automatic detection of forged fingerprints.

[0011] Furthermore, step 2) of the method includes the following steps:

[0012] 21) Use the BM3D method to denoise the fingerprint slice image collected by OCT, reduce the influence of the noise introduced in the OCT imaging process, and obtain a smooth fingerprint slice image;

[0013] 22) All fingerprint slices are adjusted in image size through bilinear interpolation and divided into image regions according to a certain window size.

[0014] Still further, step 3) of the method includes the following steps:

[0015] 31) Construct a memory-enhanced denoising autoencoder network;

[0016] 32) Use real fingerprints to train the memory-enhanced autoencoder network to obtain a trained network model.

[0017] Even further, the process of network construction in step 31) is as follows:

[0018] 311) Add Gaussian noise to the pictures input into the network model;

[0019] 312) Set the encoder, which is composed of downsampling modules, and each module is composed of a convolutional layer, a batch normalization layer and an activation layer;

[0020] 313) Set the decoder, which is composed of upsampling modules, and each module is composed of a bilinear interpolation layer, a convolutional layer, a batch normalization layer and an activation layer;

[0021] 314) Set the memory reconstruction module, which is composed of a matrix and sparse attention addressing, and the sparse attention addressing includes two parts: sparse attention calculation and threshold suppression;

[0022] 315) Set the loss function, which is composed of feature reconstruction loss and image reconstruction loss;

[0023] 316) Set the initialization of model parameters;

[0024] 317) Set the optimizer.

[0025] Preferably, step 3) further includes the following steps:

[0026] 33) Calculate the reconstruction error from the feature vectors obtained from the image patches in step 2) through the encoder and the feature vectors reconstructed through the memory, and use this error as the discrimination score.

[0027] The process of step 4) of the method is as follows: Set a threshold according to requirements such as recall rate or accuracy rate, and determine whether the fingerprint to be inspected is a real fingerprint by comparing the discrimination score obtained in step 33) with the threshold.

[0028] The beneficial effects of the present invention are mainly manifested in that the method effectively solves the problem of dependence on labeled data in supervised learning and the problem of poor robustness of manually designed anti-counterfeiting features, realizes unsupervised automatic detection of forged fingerprints, and improves the detection accuracy and generalization ability. Description of the Drawings

[0029] Figure 1 is a flowchart of an automatic forged fingerprint detection method based on OCT.

[0030] Figure 2 is a schematic diagram of intercepting a local area of a fingerprint slice.

[0031] Figure 3 is a flowchart of a memory-augmented autoencoder model. Detailed Embodiment

[0032] The present invention will be further described below in conjunction with the flowchart.

[0033] Refer to Figures 1 to 3 , an automatic forged fingerprint detection method based on optical coherence tomography, comprising the following steps:

[0034] 1) First, collect a sequence of fingerprint slices through an OCT device.

[0035] 2) Perform image preprocessing on the fingerprint slice images. The image preprocessing includes the following steps: 21) Use the BM3D method to denoise the fingerprint slice images collected by OCT, reduce the influence of noise introduced in the OCT imaging process, and obtain smooth fingerprint slice images; 22) Adjust the image size of all fingerprint slices through bilinear interpolation, and intercept a square local area on the picture, as Figure 2 shown.

[0036] 3) Refer to Figure 3, through the reconstruction of the memory-enhanced denoising autoencoder network, the reconstruction error of the latent space features is obtained. The memory-enhanced denoising autoencoder network includes the following steps: 31) Construct a memory-enhanced denoising autoencoder network; 32) Use real fingerprints to train the memory autoencoder network to obtain a trained network model; 33) Calculate the reconstruction error between the feature vector obtained by passing the image patch obtained in step 2) through the encoder and the feature vector reconstructed through the memory, and use this error as the discrimination score.

[0037] The process of network construction is as follows:

[0038] 311) For the pictures input into the network model, perform data augmentation on the real fingerprint slices input into the network model for training and add Gaussian noise.

[0039] 312) Set the encoder E. The purpose of the encoder E is to obtain the feature vector z. The encoder E can be composed of a downsampling module consisting of 5 convolutional layers.

[0040] 313) Set the decoder D. The purpose of the decoder D is to restore the original image from the reconstructed feature vector z_rec. The decoder D can be composed of 5 upsampling modules containing convolutions.

[0041] 314) Set the memory reconstruction module M, which consists of a memory and sparse attention addressing. Sparse attention addressing includes two parts: sparse attention calculation and threshold suppression. The purpose of the memory is to store the prototype features of the fingerprint slice area, and the purpose of sparse attention addressing is to calculate the importance of different memory information in the memory to the input sample. The memory can be composed of a two-dimensional matrix. Sparse attention calculation can adopt the vector similarity calculation method, and after obtaining the weights, suppress the memory information with smaller attention weights through threshold suppression.

[0042] 315) Set the loss function, which consists of the feature reconstruction loss and the image reconstruction loss. Among them, the feature reconstruction loss can be calculated by the Euclidean distance between the original feature z and the reconstructed feature z_rec, and the image reconstruction loss can be calculated by the Euclidean distance between the original image and the reconstructed image.

[0043] 316) Set the initialization of the model parameters. The parameters of the memory reconstruction module M are randomly initialized using the normal distribution.

[0044] 317) Use the Adam optimizer and set the default learning rate to 0.002.

[0045] 4) Determine whether the image patch of the fingerprint slice is a real fingerprint according to the threshold, so as to realize the automatic detection of forged fingerprints. The threshold is set according to conditions such as recall rate or accuracy. By comparing the feature loss obtained in step 33) as the discrimination score with the threshold, it is determined whether the fingerprint to be detected is a real fingerprint.

[0046] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An automatic detection method for forged fingerprints based on optical coherence tomography, characterized in that, The method includes the following steps: 1) Collect fingerprint slice images through an optical coherence tomography (OCT) device; 2) Perform image preprocessing on the fingerprint slice images; 3) Reconstruct through a memory-enhanced denoising autoencoder network to obtain the reconstruction error of the feature vectors in the latent space; 4) Determine whether the fingerprint slice image is a real fingerprint according to a threshold, so as to realize the automatic detection of forged fingerprints; The step 2) includes the following steps: 21) Use the BM3D method to denoise the fingerprint slice images collected by the OCT device, reduce the influence of the noise introduced in the OCT imaging process, and obtain smooth fingerprint slice images; 22) All fingerprint slice images are adjusted in size through bilinear interpolation, and the fingerprint slice images are divided into regions according to a certain window size to obtain multiple image blocks; The step 3) includes the following steps: 31) Construct a memory-enhanced denoising autoencoder network; 32) Use real fingerprints to train the memory-enhanced denoising autoencoder network to obtain a trained network model; The process of constructing the memory-enhanced denoising autoencoder network in the step 31) is as follows: 311) Add Gaussian noise to the pictures input into the denoising autoencoder network; 312) Set the encoder, which is composed of downsampling modules, and each downsampling module is composed of a convolutional layer, a batch normalization layer, and an activation layer; 313) Set the decoder, which is composed of upsampling modules, and each upsampling module is composed of a bilinear interpolation layer, a convolutional layer, a batch normalization layer, and an activation layer; 314) Set the memory body reconstruction module, which is composed of a matrix and sparse attention addressing, and the sparse attention addressing includes two parts: sparse attention calculation and threshold suppression; 315) Set the loss function, which is composed of feature reconstruction loss and image reconstruction loss; 316) Set the initialization of the model parameters; 317) Set the optimizer; The step 3) also includes the following steps: 33) Calculate the reconstruction error for the image blocks obtained in step 2), the feature vectors obtained through the encoder, and the feature vectors after passing through the memory body reconstruction module, and use this error as the discrimination score; The process of the step 4) is: Set the threshold according to the recall rate or accuracy condition requirement, and determine whether the fingerprint to be detected is a real fingerprint by comparing the discrimination score obtained in step 33) with the threshold.

Images