An ultrasonic guided wave fingerprint imaging and authentication method with high sensitivity and anti-counterfeiting

Abstract

The application provides an ultrasonic guided wave fingerprint imaging and authentication method with high sensitivity and anti-counterfeiting, and the steps are as follows: constructing an ultrasonic transducer for fingerprint collection and placing the ultrasonic transducer on an optical sensor to collect the acoustic field signal and the fingerprint grayscale image of the fingerprint; preprocessing the fingerprint grayscale image to establish a fingerprint velocity model; extracting the first arrival wave time domain signal segment in the acoustic field signal of the fingerprint to establish a frequency domain acoustic field matrix; establishing an offline training data set for fingerprint imaging, and obtaining a descending gradient matrix through offline training; building a post-processing model for image repair and detail matching and training the post-processing model; obtaining the fingerprint inversion image of the test example through online inversion of the descending gradient matrix; inputting the fingerprint inversion image into the trained post-processing model to obtain a fingerprint matching result; and calculating the personal authentication confidence by using the fingerprint matching result to identify the fingerprint to be authenticated. Even in the case of environmental pollution, the application can effectively identify the morphology of the epidermis on the fingertip from the acoustic field response, and has the characteristics of ultrasonic transducer screen space zero occupation, high sensitivity imaging and authentication anti-counterfeiting.

Description

Technical Field

[0001] This invention relates to the field of biometric authentication, and more particularly to an ultrasonic guided wave fingerprint imaging and authentication method. Background Technology

[0002] In recent years, biometric authentication technology has experienced rapid development in consumer electronics and public safety, not only enhancing system security but also significantly optimizing user experience. It has spawned a series of representative technologies, including fingerprint recognition, facial recognition, voice recognition, palm print recognition, iris and retinal recognition, and vein recognition. As of 2023, fingerprint authentication technology held a 34.8% share of the global biometric authentication market, demonstrating its absolute dominance, with a market value of $21.6 billion.

[0003] The combination of fingerprint scanning and authentication technologies has been widely applied in diverse electronic devices such as access control systems, mobile devices, personal computers, and wearable devices. As users increasingly demand greater comfort and universality in fingerprint authentication, and as their needs for identity security and personal privacy protection continue to evolve, fingerprint authentication technology is gradually developing towards large-scan-domain texture authentication, such as multi-fingerprint or palmprint authentication, providing an opportunity for further innovation in biometric texture authentication technology.

[0004] Under-display fingerprint scanning technology is currently systematically divided into three main categories: capacitive, optical, and ultrasonic. Among them, capacitive scanning is widely used in mobile phone unlocking and is relatively mature. This technology works by detecting the capacitance difference between the fingerprint ridges and the transducer to characterize the fingerprint pattern. However, because capacitance is difficult to effectively distinguish between the dielectric constant of the epidermis and water or oil, this technology is susceptible to interference from environmental contamination.

[0005] Optical scanning technology is commonly used in fingerprint authentication devices in public security sectors such as banking and law enforcement, offering high-resolution imaging capabilities. This technology utilizes the scattered light from fingerprint ridges for fingerprint acquisition, but the resulting image is often a 2D pattern of alternating light and dark areas, making it vulnerable to authentication attacks. Furthermore, the miniaturization of charge-coupled devices (CCDs) presents a significant technical challenge, limiting the technology's ability to achieve sufficient device thickness and thus restricting its application scenarios.

[0006] Ultrasonic scanning technology offers greater robustness. Utilizing the pulse-echo time-of-flight of sound waves, it can characterize 3D fingerprint information, demonstrating significant resistance to environmental contamination such as moisture and dust. However, because the pulse-echo involves vertical signal excitation and reception, the fingerprint scanning domain is often limited to a small area and requires a dense array of transducers. For large-area texture authentication requirements, this technology necessitates a large under-display transducer array, which not only increases cost but also negatively impacts device usability.

[0007] Traditional fingerprint scanning technology relies on signal acquisition perpendicular to the screen, which is insufficient to meet the large-scan-domain texture authentication requirements of diverse electronic devices in fields such as healthcare, banking, security, and attendance. Therefore, proposing a new method to break away from the reliance on under-display transducers has become an urgent task.

[0008] Ultrasonic guided waves are elastic waves that propagate along the surface or interior of a waveguide structure. Their propagation behavior is influenced by boundary conditions, exhibiting diverse propagation modes under different structural thicknesses and excitation frequencies. This dispersion phenomenon is extremely sensitive to minute changes in the structural surface, allowing for the differentiation of different waveguide materials by analyzing the acoustic field response characteristics. By arranging a series of transducers in an array to surround the target region and combining this with tomographic imaging to detect changes in the acoustic field, structural changes within the scanning domain can be characterized using the transducer network. This method achieves target imaging without requiring the movement of individual transducers for scanning.

[0009] For example, the invention patent with application number 202311018944.7 relates to the field of ultrasonic fingerprint recognition technology, and particularly to an ultrasonic fingerprint imaging method, fingerprint recognition device, and electronic device. The method includes the following steps: setting ultrasonic transducers to be distributed in a uniform rectangular array on a sensor; dividing each n×n ultrasonic transducer array into a group; emitting multiple ultrasonic waves from the ultrasonic transducers to the object; receiving and outputting multiple ultrasonic signals reflected from the object; based on the biomimetic principle of image convolution, adjusting the amplitude of each pixel in the ultrasonic transducer array individually based on the values ​​of each point in the convolution kernel; based on the principle of beamforming, adjusting the phase of each pixel in the ultrasonic transducer array individually; focusing the acoustic wave intensity of surrounding pixels onto the central pixel using beamforming technology; and acquiring the fingerprint image of the object at the central pixel. However, the above application uses an under-display ultrasonic transducer array, which characterizes the fingerprint morphology by measuring the flight time and intensity of the pulse echo to the fingerprint. Although it has the inherent robustness of ultrasonic technology in resisting environmental pollution, its scanning domain is basically the same size as the transducer array, which limits its application in fingerprint characterization in a large scanning domain. It also has certain limitations in fingerprint authentication at any location, multiple fingerprints, and palm print authentication. Summary of the Invention

[0010] To address the technical problems of existing fingerprint scanning technologies, which rely on complex under-display devices leading to a limited scanning domain, and the susceptibility of capacitive and optical fingerprints to deception by fake textures and interference from environmental pollution, this invention proposes a highly sensitive and anti-counterfeiting ultrasonic guided wave fingerprint imaging and authentication method. It utilizes signal acquisition in the plane direction of the plate to detect changes in the acoustic field within the scanning domain, and employs ultrasonic guided wave tomography based on supervised descent to characterize sub-millimeter-level textures.

[0011] To achieve the above objectives, the technical solution of the present invention is as follows: a highly sensitive and anti-counterfeiting ultrasonic guided wave fingerprint imaging and authentication method, comprising the following steps:

[0012] S1: Construct an ultrasonic transducer for fingerprint acquisition and place the ultrasonic transducer on an optical sensor. The ultrasonic transducer acquires the acoustic field signal of the fingerprint, and the optical sensor acquires the grayscale image of the fingerprint.

[0013] S2: Preprocess the fingerprint grayscale image and calculate the phase velocity range to characterize the fingerprint based on the global matrix method. Establish a fingerprint velocity model based on the preprocessed fingerprint grayscale image.

[0014] S3: Extract the first arrival wave time domain signal segment of the phase velocity interval corresponding to the pattern in the acoustic field signal of the fingerprint, extract the frequency domain energy flow amplitude to obtain the spectrum amplitude, and establish the frequency domain acoustic field matrix;

[0015] S4: Build an offline training dataset for fingerprint imaging using the methods in steps S2 and S3, obtain the descent gradient matrix through offline training; build and train a post-processing model for image inpainting and detail matching.

[0016] S5: Acquire the acoustic field signal of the fingerprint to be authenticated through an ultrasonic transducer, acquire the grayscale image of the fingerprint to be authenticated through an optical sensor, obtain the fingerprint velocity model and frequency domain acoustic field matrix of the test case using the methods in steps S2 and S3, obtain the fingerprint inversion image of the test case through online inversion of the descent gradient matrix; input the fingerprint inversion image into the trained post-processing model to obtain the fingerprint matching result.

[0017] S6: Calculate personal authentication confidence using fingerprint matching results, and identify the fingerprint to be authenticated using personal authentication confidence.

[0018] Preferably, the ultrasonic transducer includes a plurality of ultrasonic transducer elements arranged in a ring. Each ultrasonic transducer element includes a piezoelectric ceramic sheet group. A positive copper foil and a negative copper foil are respectively disposed on the upper and lower sides of the piezoelectric ceramic sheet group. A glass plate is disposed on the lower side of the negative copper foil and the upper side of the positive copper foil. Both the positive and negative copper foils are connected to the ultrasonic instrument. An insulating layer is laid between the positive and negative copper foils, and a sound field boundary absorption layer is disposed at the outermost edge between the two glass plates.

[0019] Preferably, the piezoelectric ceramic sheet assembly consists of at least one circular PZT-5H sheet, with the positive and negative copper foils respectively fixedly connected to the PZT-5H sheet by conductive silver paste, and the negative copper foil fixed to the glass plate below the negative copper foil by an epoxy resin adhesive.

[0020] Preferably, the PZT-5H sheet is made by cutting and polishing PZT-5H blank, and both the positive and negative copper foils are obtained by laser cutting;

[0021] Both the positive and negative copper foils include circular endpoints with the same diameter as the PZT-5H sheet. Leads are provided on the circular endpoints. The ends of the leads of the negative copper foil are connected to the annular copper foil, which is grounded. The outermost edge of the negative copper foil is directly connected to a square copper foil. The four corners of the square copper foil are divided into straight copper foils and right-angled copper foils by breakpoints. The straight copper foils are folded 90° outwards, and the copper-clad surface is flipped to face downwards on the same side as the copper-clad surface of the positive copper foil. The right-angled copper foils are reserved as debugging contact points.

[0022] The positive electrode copper foil is divided into four parts, with the copper-clad surface of each part facing down; each part of the positive electrode copper foil is fixed with a polydimethylsiloxane film; the outermost edge of the glass plate is surrounded by modeling clay to serve as a sound field boundary absorption layer;

[0023] A transparent polydimethylsiloxane film is laid between the surface glass of the optical sensor and the glass plate of the ultrasonic transducer.

[0024] Preferably, the method for preprocessing the fingerprint grayscale image is as follows: fingerprint tracking and preliminary feature extraction are performed to obtain preliminary features of the fingerprint image; target regions in the preliminary features of the fingerprint image are extracted by threshold control; the extracted target regions are converted into low dynamic range images using a bilateral filtering algorithm; and connected regions in fingerprint details are eliminated by a spatial filtering algorithm with convolution kernels to obtain fingerprint ridge images.

[0025] Preferably, the method for establishing the fingerprint velocity model is as follows: using the thickness of the fingertip epithelium, the height difference between the ridges and valleys, the epithelial density, Young's modulus, and Poisson's ratio as inputs, the phase velocity dispersion curves before and after the fingertip ridge is attached to the glass plate of the ultrasonic transducer are calculated using the global matrix method. At the same frequency, the A0 mode, which has a shorter wavelength and carries more acoustic field information, is selected. Compared with the glass structure of the glass plate, the material properties of the fingertip epithelium are closer to those of a fluid. Fluid-structure interaction theory shows that after the ridge is attached to the glass, the acoustic field energy flow is mainly concentrated in the adjacent modes below the phase velocity dispersion curve of the glass plate. At the center frequency of the ultrasonic transducer, the attachment of the ridge to the glass plate causes the phase velocity of the A0 mode to decrease from 2.91 mm / μs to 2.55 mm / μs, resulting in a phase velocity range of [2.55, 2.91] mm / μs.

[0026] The fingerprint ridge image is mapped to the phase velocity range [2.55, 2.91] mm / μs to obtain the fingerprint velocity model.

[0027] Preferably, the method for extracting the first arrival time domain signal segment of the phase velocity interval corresponding to the fingerprint acoustic field signal is as follows: the group velocity dispersion curve before and after the ridge is attached to the glass plate is calculated using the conversion relationship between phase velocity and group velocity. At the center frequency of the ultrasonic transducer, there is an obvious mode separation phenomenon between the A0 mode and other modes. The time domain signal of the A0 mode is extracted from the time domain signal of the acquired acoustic field signal using the group velocity difference and window function.

[0028] The method for extracting the frequency domain energy flow amplitude to obtain the spectral amplitude is as follows: the time domain signal segment of each ultrasonic transducer array element is transformed into a frequency domain signal by frequency domain transformation, the frequency domain signal is smoothed, the peak value of the frequency domain signal spectrum after smoothing is extracted to obtain the spectral amplitude, the extracted spectral amplitude is written into the corresponding point of the frequency domain sound field matrix, and the frequency domain sound field matrix is ​​mapped to the interval [1,2].

[0029] Preferably, the supervised descent method used in the offline training is:

[0030] R=(Δs T Δs+β 2 I) -1 ·(Δs T Δm);

[0031] In the formula, R is the descent gradient matrix obtained through offline training, Δs = s0 - s1 is the difference in the sound field matrix, where s0 and s1 represent the frequency domain sound field matrix before and after fingertip pressing, respectively; β is the regularization weight, Δm = m0 - m1 is the difference in the fingerprint velocity model, where m0 and m1 represent the fingerprint velocity model before and after fingertip pressing, respectively; the velocity sound field matrix s and the fingerprint velocity model m are both data matrices containing all examples;

[0032] Online inversion is performed using the descent gradient matrix R as a test case to obtain the imaging result M of the test case. test =m0+Δs test ·R;Δs test This is the frequency domain sound field matrix corresponding to the test case.

[0033] Preferably, the method for fingerprint tracking and preliminary feature extraction is as follows: in the fingerprint grayscale image, track the point with the largest grayscale value, take the point with the largest grayscale value as the center, extract the grayscale image within a range of 36mm×36mm around it, invert the grayscale image, and extract the range with grayscale values ​​below 110; the threshold control method is to discretize the inverted grayscale image with grid spacing, and retain the part with an average grayscale value below 80 within the grid.

[0034] The smoothing process employs a Gaussian weighted moving average filtering algorithm.

[0035] The method for establishing the offline training dataset is as follows: collect sound field signals and fingerprint grayscale images of the fingertip epithelium under different dry conditions, at different locations, and with different pressures. The fingerprint velocity model and sound field signal obtained for each fingertip press are combined into a calculation example through the methods in steps S2 and S3.

[0036] Preferably, the post-processing model includes a DLCNN model for image restoration and a Mask R-CNN model for fingerprint matching. The DLCNN model adopts a small-scale, deep architecture, with all convolutional layers using a 3×3 dimension, and the network depth of the DLCNN model is 20 layers. The Mask R-CNN model includes a ResNet101 feature extraction network, a feature pyramid network, a region proposal network, a region of interest arrangement, and a fully convolutional network connected in sequence. The restored fingerprint grayscale image is obtained using the DLCNN model, and after being converted into a binary form of the fingerprint image by an even-symmetric Gabor filter, it is input into the Mask R-CNN model. The restored fingerprint grayscale image is then processed by the Mask R-CNN model to obtain the fingerprint matching result image.

[0037] The training dataset of the DLCNN model includes grayscale images of fingerprints from the artifact removal part and grayscale images of natural landscape images from the image denoising part. Each case of the artifact removal part includes an image of fingerprint inversion with artifacts and a corresponding clean fingerprint image; each case of the image denoising part includes a natural landscape image with Gaussian noise and a corresponding clean image.

[0038] The training dataset for the Mask R-CNN model includes binarized grayscale images of fingerprints from multiple registered fingertips acquired by an optical sensor, with fingerprint markers attached.

[0039] A wide-range marking method covering multiple fingerprint details was adopted, marking 8 wide-range fingerprint details;

[0040] The method for calculating personal authentication confidence is as follows: the uniqueness of fingerprint details of the registered fingertip is divided into three levels. Level 1 fingerprint details have the highest uniqueness and are selected as key features representing personal information. Level 2 fingerprint details have the next highest uniqueness, while Level 3 fingerprint details have relatively weaker uniqueness. Therefore, the personal authentication confidence is:

[0041]

[0042] In the formula, ξ k The compensation factor is an empirical value related to the labeled fingerprint details, which varies with the number of fingerprint details k∈[0,8] identified by the Mask R-CNN model, and b is the weight of level 1 to 3.

[0043] Compared with existing technologies, the advantages of this invention are as follows: An ultrasonic transducer array is arranged on the upper surface at the edge of the large scanning domain, so that the ultrasonic transducers do not occupy under-screen space; signal acquisition in the plane direction of the plate is used to detect changes in the acoustic field within the scanning domain, and ultrasonic guided wave tomography based on supervised descent is used to characterize sub-millimeter-level textures. An advanced post-processing model is introduced for image restoration and detail matching, improving the reconstructed fingerprint detail morphology and providing anti-counterfeiting features for personal authentication.

[0044] The ultrasonic guided wave fingerprint imaging and authentication method proposed in this invention features zero space occupation under the transducer screen, high-sensitivity imaging, and authentication anti-counterfeiting capabilities. It uses signals acquired in the plane of the plate and leverages the Lamb wave dispersion effect generated when a fingertip presses against the glass plate to characterize sub-millimeter-level texture details. It learns the nonlinear mapping relationship between sound field changes and fingerprint morphology changes, enabling effective identification and recognition of the fingertip epithelial layer morphology from the sound field response even under polluted environmental conditions. This method not only provides an effective solution for large-scan-domain fingerprint characterization but also foreshadows its development prospects in multi-fingerprint characterization and palmprint authentication technologies. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of the present invention.

[0047] Figure 2 This is a schematic diagram of fingerprint acquisition operation according to an embodiment of the present invention.

[0048] Figure 3 This invention relates to the electrical interconnection method of the PZT-5H ultrasonic transducer array elements.

[0049] Figure 4 This is a schematic diagram of the fingerprint acquisition acoustic transducer developed in this invention.

[0050] Figure 5 This invention relates to a method for establishing a fingerprint speed model.

[0051] Figure 6 This is a side view of the sound field propagation when the fingertip presses on the glass plate as developed in this invention.

[0052] Figure 7The figures are Lamb wave phase velocity dispersion curves before and after fingertip pressing according to an embodiment of the present invention. (a) is a Lamb wave phase velocity dispersion curve in the frequency range of 0-7MHz, and (b) is an enlarged view of the dashed box in (a).

[0053] Figure 8 This is a schematic diagram of obtaining the frequency domain sound field matrix proposed in this invention, wherein (a) is the extracted A0 mode time domain signal, (b) is the A0 mode frequency domain energy flux amplitude, and (c) is the frequency domain sound field matrix.

[0054] Figure 9 The figures show the Lamb group velocity dispersion curves before and after fingertip pressing according to an embodiment of the present invention, wherein (a) is the Lamb group velocity dispersion curve in the frequency range of 0-7MHz, and (b) is an enlarged view of the dashed box in (a).

[0055] Figure 10 This is a flowchart of the image post-processing model proposed in this invention.

[0056] Figure 11 This is a schematic diagram illustrating an imaging and authentication of a clean registered fingertip, and the authentication attack result of a clean unregistered fingertip, as described in an embodiment of the present invention.

[0057] Figure 12 This is a schematic diagram illustrating the imaging and authentication of a registered fingertip with moisture and a registered fingertip with dust, as described in an embodiment of the present invention.

[0058] In the diagram, 1 is the piezoelectric ceramic sheet assembly, 2 is the positive electrode copper foil, 3 is the negative electrode copper foil, 4 is the glass plate, 5 is the epoxy resin binder, 6 is the conductive silver paste, 7 is the polydimethylsiloxane film, 8 is the insulating layer, 9 is the sound field boundary absorption layer, and 10 is the debugging contact point. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] like Figure 1As shown, a high-sensitivity and anti-counterfeiting ultrasonic guided wave fingerprint imaging and authentication method is proposed. Targeting three main objectives—zero space occupancy under the transducer screen, high-sensitivity imaging, and anti-counterfeiting authentication—a fingerprint acquisition ultrasonic transducer with a φ100mm scanning domain was developed, and a complete fingerprint acquisition system was built. This system combines the data acquisition and signal processing methods required for fingerprint imaging and authentication. By obtaining the frequency domain acoustic field difference and velocity model difference caused by fingertip pressure, an offline training dataset for fingerprint imaging is established. Supervised descent is used to learn the nonlinear mapping relationship between the two differences, which is then used for rapid imaging of test case fingerprints. In the image post-processing model, DLCNN (Deep Learning Convolutional Neural Network) and Mask R-CNN (Mask Region-based Convolutional Neural Network) models are employed. The DLCNN model is used for artifact removal and image denoising, improving the reconstruction of fingerprint details and significantly enhancing reconstruction accuracy. The Mask R-CNN model is used for detail matching, providing anti-counterfeiting authentication. The specific steps are as follows:

[0061] S1: Construct a fingerprint acquisition ultrasonic transducer and place it on an optical sensor. The fingerprint acquisition ultrasonic transducer acquires the acoustic field signal of the fingerprint, and the optical sensor acquires the grayscale image of the fingerprint.

[0062] like Figure 2 As shown, the ultrasonic guided wave fingerprint imaging method detects changes in the acoustic field within the scanning domain by acquiring signals along the plane of the plate, which is used to characterize sub-millimeter-level textures such as fingerprints. To meet the verification requirements of ultrasonic guided wave fingerprint imaging, a fingerprint acquisition transducer device with a φ100mm scanning domain was designed and developed in this embodiment. The range within the ultrasonic transducer array is called the scanning domain, and the diameter of the scanning domain is determined by the size and number of ultrasonic transducers. The ultrasonic transducers are obtained by grinding PZT-5H material blanks. To ensure effective displacement in the z-direction (perpendicular to the glass plate), they cannot be made too small; therefore, the diameter of the transducer array elements processed in this embodiment is 2mm. Figure 3 As shown; regarding the number of transducers, to ensure sufficient fingerprint inversion image resolution, the number of transducers cannot be too small. This embodiment uses 128; if there are too many, the computational load will be too large. Therefore, the diameter of the scanning domain is not absolute; it can be made smaller or larger, and the attenuation of the guided wave field needs to be considered. Current experimental and simulation results show that it is fine even for a laptop screen.

[0063] Specifically, such as Figure 3As shown, the fingerprint acquisition ultrasonic transducer includes several ultrasonic transducer elements arranged in a circular pattern, such as... Figure 4 As shown, the ultrasonic transducer array element includes a circular piezoelectric ceramic sheet group 1. A positive copper foil 2 and a negative copper foil 3 are respectively disposed on the upper and lower sides of the piezoelectric ceramic sheet group 1. A glass plate 4 is disposed on the lower side of the negative copper foil 3 and the upper side of the positive copper foil 2. A sound field boundary absorption layer 9 is disposed at the outermost edge between the two glass plates. The positive copper foil 2 is connected to the ultrasonic instrument via an interconnect module, and the negative copper foil 3 is connected to the ultrasonic instrument's grounding terminal. The interconnect module serves as a signal acquisition and data transfer interface, enabling the generation of excitation signals for the ultrasonic transducer array element and the acquisition of guided wave signals.

[0064] Among them, the piezoelectric ceramic sheet group 1 consists of at least one PZT-5H sheet. The PZT-5H sheet is made by cutting and polishing a PZT-5H blank with a center frequency of 5MHz. It is a circular sheet with a diameter and thickness of approximately 2mm and 0.5mm, respectively.

[0065] Both the positive and negative copper foils 2 and 3 are obtained by laser cutting. Each foil includes a circular end with a diameter of 2 mm, and a 0.7 mm wide lead is provided on the circular end. Each piezoelectric ceramic sheet group 1 corresponds to one positive copper foil 2 and one negative copper foil 3. The positive and negative copper foils 2 and 3 are fixedly connected to the PZT-5H sheet using conductive silver paste 6 without affecting conductivity. The negative copper foil 3 is fixed to the glass plate below it using an epoxy resin adhesive 5. This adhesive acts as a sound field coupling agent, ensuring sound field transmission between the ultrasonic transducer and the glass plate. After the transducers are electrically interconnected, the center frequency decreases due to the overall thickness increase. In this embodiment, impedance analysis determined the average center frequency of the 128-channel transducer to be approximately 4.29 MHz.

[0066] In this embodiment, a transducer array composed of 128 PZT-5H thin sheets is used. The channel numbers are assigned starting from the ultrasonic transducer element in the positive x-axis direction and increasing counterclockwise. The PZT-5H thin sheets are arranged in a ring. During use, multiple ultrasonic transducer elements transmit excitation signals in numerical order, and all elements simultaneously acquire guided wave signals. The glass plate measures 120mm × 120mm × 0.55mm and is Schott B270 soda-lime glass with a density ρ of 2.56g / cm³. 3The Young's modulus E is 71.10 GPa, and the Poisson's ratio ν is 0.22. The leads of the negative electrode copper foil 3 are all connected to a ring-shaped copper foil with a radius of 60 mm, and the ring-shaped copper foil is grounded. At the outermost edge of the negative electrode copper foil 3, it is directly connected to a square copper foil with a side length of 120 mm. The four corners of the square copper foil are divided into straight copper foil and right-angled copper foil. The straight copper foil is folded 90° outwards, with its copper-clad surface facing downwards, on the same side as the copper-clad surface of the positive electrode copper foil, and connected to the interconnect module together. The right-angled copper foil is reserved as the device's debugging contact point 10.

[0067] The positive electrode copper foil is divided into four parts, each responsible for the electrical interconnection of 32 channels. These four parts connect to the ultrasonic instrument from four directions, with the copper surface facing downwards. Each part, along with a straight copper foil of a flipped negative electrode copper foil 3, is connected to the interconnect module and ultimately to the ultrasonic instrument. To ensure the stability and flexibility of the positive electrode copper foil, each part is fixed with a polydimethylsiloxane film 7, forming a bus. An insulating layer 8 is laid between the positive and negative electrode copper foils. The outermost edge of the glass plate is surrounded by modeling clay, acting as a sound field boundary absorbing layer 9.

[0068] The key to achieving fingerprint imaging lies in offline training to learn the nonlinear mapping relationship between fingerprint morphology and sound field changes when a fingertip is pressed. Therefore, in addition to an ultrasonic instrument, a device capable of precisely capturing fingerprint morphology is also required. In this embodiment of the invention, the fingerprint acquisition ultrasonic transducer is placed above an optical sensor. This optical sensor is an EPSON V600, with image storage and resolution set to grayscale and 1200 dpi, respectively. To reduce the potential interference of the optical sensor's surface glass on the sound field of the fingerprint acquisition ultrasonic transducer device, a transparent polydimethylsiloxane film is laid between the two.

[0069] An optical sensor is used to capture the fingerprint pattern when the fingertip is pressed, and the Lamb wave dispersion effect is used to map it to the corresponding phase velocity range to obtain a fingerprint velocity model. An ultrasonic transducer for fingerprint acquisition is used to capture the sound field signal when the fingertip is pressed. The change in the sound field signal caused by fingertip pressing compared to the unpressed state is called the sound field change. The fingerprint velocity model and sound field signal obtained for each fingertip press constitute a calculation example.

[0070] S2: Preprocess the fingerprint grayscale image and calculate the phase velocity range to characterize the fingerprint based on the global matrix method to establish a fingerprint velocity model.

[0071] like Figure 5 As shown, the fingerprint speed model serves as a reference standard for fingerprint morphology, and the fingerprint grayscale image is acquired by an optical sensor placed below the fingerprint acquisition transducer device.

[0072] The fingerprint grayscale image obtained in step S1 is preprocessed. The fingerprint image is a grayscale image with a grayscale range of 0-255. In fingerprint image processing, firstly, fingerprint tracking and preliminary feature extraction are performed to obtain preliminary features of the fingerprint image. The specific method for fingerprint tracking and preliminary feature extraction is as follows: In the fingerprint grayscale image, within the range corresponding to the fingerprint acquisition ultrasonic transducer, the point with the largest grayscale value is tracked, i.e., the brightest point of the fingerprint. Taking the point with the largest grayscale value as the center, the grayscale image within a range of 36mm×36mm around it is extracted. This grayscale image is inverted to convert the fingerprint pixels from white to black, and the range with grayscale values ​​below 110 is extracted. Since the white part in the grayscale image after inversion is basically glass, the near-pure white part in the image is omitted. Next, the target area is extracted from the preliminary features of the grayscale image after inversion through threshold control. The grayscale image after inversion is discretized with a grid spacing of 0.4mm, and the part with an average grayscale value below 80 within the grid is retained to resist the interference of dust and impurities on the fingerprint image and further narrow down the range of valid fingerprints. To reduce fingerprint image complexity and contrast while maintaining image quality, a bilateral filtering algorithm is used to convert the extracted target region image into a low dynamic range (LDR) image. Finally, a spatial filtering algorithm with convolution kernels is introduced. This algorithm primarily eliminates connected regions in fingerprint details, transforming the LDR image into a clear fingerprint ridge image, thereby reducing local contrast differences caused by uneven pressure applied by the fingerprint tip.

[0073] like Figure 6 As shown, fingerprints are composed of ridges and valleys. When a fingertip is pressed against a glass plate, the fingerprint ridges adhere to the glass material, while the fingerprint valleys do not. The adhesion between the ridges and the glass plate forms a double-layer structure, and the dispersion effect of this structure is completely different from that of a single-layer glass plate. Figure 6 The ultrasonic transducer array elements emit different excitations, and the waveform changes of the guided waves generated by the excitations before and after encountering the fingertip press are collected to obtain the sound field signal.

[0074] In this embodiment, the thickness of the fingertip epithelium is defined as 1.5 mm, the height difference between the ridges and valleys is 0.25 mm, and the density ρ is 1.02 g / cm³. 3 The Young's modulus E is 2.19 GPa, and the Poisson's ratio ν is 0.4. Using the thickness of the fingertip epithelium, the height difference between the ridges and valleys, the epithelial density, Young's modulus, and Poisson's ratio as inputs, the phase velocity dispersion curves before and after the ridges are attached to the glass plate are calculated using the global matrix method. The results are as follows: Figure 7As shown in the figure. The gray solid curve represents the dispersion curve of the glass plate before fingertip pressure, and the black solid curve represents the dispersion curve of the double-layer material formed by the ridge and glass bonding after fingertip pressure. The fingerprint acquisition ultrasonic transducer operates at a frequency of 4.29MHz. In the figure, c... p The phase velocity is represented by A and S in the diagram. A and S represent the antisymmetric and symmetric modes of the Lamb wave, respectively. Within the 0-7MHz frequency range, the A and S modes appearing from left to right are named A0, A1, S0, and S1 modes, respectively. Theoretically, the shorter wavelength A0 mode carries more sound field information at the same frequency. Compared to the glass structure of the glass plate, the material properties of the fingertip epithelium are closer to those of a fluid. Fluid-structure interaction theory indicates that after the ridge is attached to the glass, the sound field energy flow is mainly concentrated in the adjacent modes below the phase velocity dispersion curve of the glass plate. At the operating frequency of 4.29MHz, the attachment of the ridge to the glass plate causes the phase velocity of the A0 mode to decrease from 2.91mm / μs to 2.55mm / μs, a decrease of 12.37% in sound velocity.

[0075] By mapping the fingerprint ridge image to the phase velocity range [2.55, 2.91] mm / μs, a fingerprint velocity model is obtained, which can represent the thickness of the fingertip pressing the glass plate as a phase velocity, that is, converting 3D information into a 2D image, thereby improving the computational speed of the fingerprint imaging method.

[0076] S3: Extract the first arrival wave time domain signal segment of the phase velocity interval corresponding to the fingerprint sound field signal, extract the frequency domain energy flow amplitude to obtain the spectrum amplitude, and establish the frequency domain sound field matrix.

[0077] In this embodiment, the excitation signal is a sine wave signal with a center frequency of 4.29MHz, an amplitude of 10V, and 10 cycles. Figure 8 The time-domain signal between two transducers on the diameter of the ultrasonic transducer array is shown. In the A0 mode frequency domain energy flow amplitude extraction step, the smoothing process adopts the Gaussian weighted moving average filtering algorithm.

[0078] Using the conversion relationship between phase velocity and group velocity, the group velocity dispersion curves before and after the ridge-bonded glass plate are calculated, such as... Figure 9 As shown in the figure, c g The group velocity is represented by the group velocity. At the operating frequency of 4.29MHz, the group velocity of mode A0 is 3.39mm / μs. There is a clear mode separation between mode A0 and modes A1 and S0, indicating that mode A0 is easy to extract and has little overlap with the time-domain signals of modes A1 and S0. Utilizing the group velocity difference and a window function, the time-domain signal of mode A0 is extracted from the acquired sound field signal.

[0079] To further improve computational speed, the time-domain signal acquired by the ultrasonic instrument was processed into a frequency-domain signal. The extracted A0 mode time-domain signal was converted into a frequency-domain signal using a Fast Fourier Transform. To reduce the complexity of the frequency-domain signal, a Gaussian weighted moving average filtering algorithm was used to smooth the spectrum. The smoothed spectrum showed that pressing the fingertip caused the relative amplitude of the sound field energy flow between the two transducers on the diameter to decrease from 3.47 to 2.35, an amplitude of 32.28%. This indicates that the adhesion between the fingerprint ridge and the glass plate has a significant impact on the sound field, and the sound field effect is theoretically proportional to the imaging quality. The peak values ​​of the spectrum were extracted, and the smoothed spectrum amplitudes were extracted and written into the corresponding points of a 128×128 frequency-domain sound field matrix. The frequency-domain sound field matrix was mapped to the interval [1,2] to avoid non-zero values ​​caused by the presence of 0 in the imaging calculation.

[0080] The ultrasonic transducer array employs a method of sequential excitation of individual elements in a counter-clockwise direction and simultaneous reception of the entire array. Therefore, using the aforementioned frequency domain energy flow amplitude extraction method, a frequency domain sound field matrix with dimensions of 128×128 can be established. To avoid non-numerical problems caused by the presence of zero values ​​during the calculation process, the sound field energy flow amplitude is mapped to the interval [1,2].

[0081] The pure black grayscale image of the optical sensor when it is unloaded and the sound field collected by the fingerprint acquisition ultrasonic transducer when it is unloaded together constitute the baseline signal for obtaining the velocity model difference and the frequency domain sound field difference.

[0082] S4: Establish an offline training dataset for fingerprint imaging and obtain the descent gradient matrix through offline training; the fingerprint to be tested obtains the fingerprint velocity model and frequency domain acoustic field matrix of the test case through steps S1-S3, and obtains the fingerprint inversion image of the test case through online inversion of the descent gradient matrix.

[0083] To enhance the breadth and diversity of the dataset, an offline training dataset was established in this embodiment. The dataset contains 2048 imaging examples of fingertip epithelial images under different dry conditions, at different locations, and with different pressure applied. Each example includes a corresponding fingerprint velocity model and a frequency domain acoustic field matrix.

[0084] Offline fingerprint imaging training was performed on an HP workstation equipped with 32 Intel Xeon Gold 6230 CPUs, using MATLAB R2020a as the runtime environment. The core formula of the supervised descent method used in offline training is as follows:

[0085] R=(Δs T Δs+β 2 I) -1 ·(Δs T Δm);

[0086] In the formula, R is the descent gradient matrix obtained through offline training, Δs = s0 - s1 is the difference in the sound field matrix, and s0 and s1 represent the frequency domain sound field matrices before and after fingertip pressing, respectively. β is a user-defined regularization weight; in this embodiment, the value of this variable is 1 × 10⁻¹⁰. -3 I is the identity matrix, and Δm = m0 - m1 is the fingerprint velocity model difference, where m0 and m1 represent the fingerprint velocity model before and after fingertip pressure, respectively. To improve the computational efficiency of offline training, the velocity sound field matrix s and the fingerprint velocity model m are both large matrices containing all examples. This allows the supervised descent method to output the descent gradient matrix R obtained through offline training in one iteration, i.e., after all examples have been trained once.

[0087] The fingerprint imaging method of this invention concentrates most of the computation in the offline training stage. The descent gradient matrix R obtained through offline training is directly used for online inversion of test cases.

[0088] M test =m0+Δs test ·R;

[0089] In the formula, M test For the imaging results of the test case, Δs test This is the frequency domain sound field matrix corresponding to the test case.

[0090] S5: Build and train a post-processing model for image restoration and detail matching. The fingerprint inversion image of the test case is used to obtain the fingerprint matching result through the trained post-processing model.

[0091] The post-processing models include a DLCNN model for image inpainting and a Mask R-CNN model for fingerprint matching, such as... Figure 10As shown, the model includes a Deep Learning Convolutional Neural Network (DLCNN) model for image inpainting and a Mask R-CNN model for fingerprint matching, aiming to improve the quality of reconstructed fingerprint images and enhance authentication anti-counterfeiting capabilities. The DLCNN model employs a small-scale, deep architecture, with all convolutional layers using a 3×3 dimension and a network depth of 20 layers. Linear rectified units (RCUs) are placed after convolutional layers 1-19, and batch normalization is applied between convolutional layers 2-19 and the RCUs. This architecture is suitable for image processing. The RCU is an activation function that primarily increases the non-linearity of the network, enabling it to learn more complex features and improving training efficiency. Batch normalization is a technique used to improve training speed and stability. By normalizing each mini-batch of data, the output distribution of the intermediate layers becomes more stable, reducing internal covariate shift issues. The fingerprint inversion images of the test cases are converted to grayscale and cropped before being input into the DLCNN model's input layer. This cropping of the input images is to adapt to the network structure, improve processing efficiency, and enhance the inpainting effect. Its main functions include adapting to the network input size, reducing computational load, avoiding boundary effects, and improving the model's generalization ability. After cropping, the image is sequentially input into each convolutional layer of the intermediate layers through the image input layer. The output of the intermediate layers is then used for regression analysis to obtain the restored fingerprint grayscale image.

[0092] In this embodiment, the training dataset for the DLCNN model is divided into two parts: one part is used to train artifact removal capability, and the other part is used to train image denoising capability. Each example in the training dataset includes a noisy image containing artifacts or noise and a clean image. The training dataset includes grayscale images of fingerprints and natural landscape images, used to train artifact removal performance and image denoising performance. These two datasets contain 13195 and 150 examples respectively. Each example contains two grayscale images: for artifact removal, one fingerprint inversion image containing artifacts and one corresponding clean fingerprint image; for image denoising, one natural landscape image containing Gaussian noise and one corresponding clean image. Artifact removal is the primary function, and denoising is a secondary function, so there are fewer denoising examples. Moreover, the denoising training uses complex natural landscape images, which can improve the comprehensiveness of denoising as much as possible with fewer examples. The training dataset is divided into training and test sets in an 8:2 ratio. The training process was performed in the Deep Learning Toolbox of MATLAB R2020a and accelerated using an NVIDIA GeForce RTX 2080Ti GPU. The optimizer used was SGDM (stochastic gradient descent momentum), with the gradient threshold and momentum parameter set to 5 × 10⁻⁶. -3The batch size, maximum number of iterations, initial learning rate, and L2 tradeoff parameter were set to 32, 30, 0.1, and 0.1 × 10⁻⁹, respectively. -3 .

[0093] The restored fingerprint grayscale image is obtained using a DLCNN model. An even-symmetric Gabor filter is then used to transform the restored fingerprint grayscale image into a binary form of the fingerprint image. This binary form is then input into a Mask R-CNN model to enhance the ridge and valley information of the fingerprint. The restored fingerprint grayscale image is then used by the Mask R-CNN model to obtain the fingerprint matching result image.

[0094] The Mask R-CNN model consists of five sequentially connected parts: a ResNet101 feature extraction network, a feature pyramid network, a region proposal network, a region of interest (ROI) arrangement network, and a fully convolutional network. The ridge ends and bifurcations of a fingerprint are key details representing personal information. To incorporate the relative positional information of multiple fingerprint details, a large-scale labeling method covering multiple fingerprint details was adopted, labeling a total of eight large-scale fingerprint details. Fingerprint labeling was done manually, and different registered fingerprints yielded different labeling results. The labeled detail image is shown below. Figure 10 The fingerprint detail markers at the far right and top are generated using the Labelme 5.2.1 plugin. The output is a .json file containing the annotation information, which is directly input into the Mask R-CNN model for training. In other words, the binary dataset is labeled, converted to a .json file, and then training begins.

[0095] In this embodiment, the right fingertip of the second inventor is used as the registered fingertip. The training dataset for the Mask R-CNN model consists of the fingerprint image of the registered fingertip and its detail markers. Through training, the Mask R-CNN model can identify the classification, regression analysis, and masking of these details, thereby achieving detail matching of the test case fingerprint and outputting its fingerprint matching information and image.

[0096] The training dataset for the Mask R-CNN model consists of binarized grayscale fingerprint images from 4680 optical sensors, each with fingerprint markers. The training dataset is divided into training and test sets in a 9:1 ratio. Training was performed using Tensorflow 2.6.0 and PyCharm as the code editor, accelerated by an NVIDIA GeForce RTX 2080Ti GPU. The batch size and initial learning rate were set to 16 and 0.02, respectively.

[0097] S6: Establish a personal authentication confidence algorithm, use fingerprint matching results to calculate personal authentication confidence, and use it for fingerprint recognition.

[0098] Among the eight widely marked fingerprint details, those with simple and symmetrical shapes exhibited higher uniqueness. Therefore, the uniqueness of the fingerprint details of the registered fingertip was divided into three levels, such as... Figure 10 As shown. In dense fingerprint patterns, simple and symmetrical details possess higher uniqueness. Level 1 fingerprint details have the highest uniqueness and are selected as key features representing personal information; Level 2 fingerprint details have the next highest uniqueness, while Level 3 fingerprint details have relatively weaker uniqueness. The formula for calculating personal authentication confidence for this registered fingerprint tip is:

[0099]

[0100] In the formula, c f This represents the credibility of an individual's authentication. k The compensation factor is an empirical value related to the labeled fingerprint details, varying with the number of fingerprint details k∈[0,8] recognized by the Mask R-CNN model. The values ​​of ξ0 to ξ8 are 0, 3, 3, 15 / 7, 5 / 3, 10 / 7, 5 / 4, 10 / 9, and 1, respectively. b is the weight, with weights b corresponding to fingerprint details of levels 1 to 3 being 1 / 6, 2 / 15, and 1 / 10, respectively. When all 8 labeled large-scale fingerprint details are recognized, the personal authentication confidence c... f The value is 0.99. Compensation factor ξ k Both the value of b and the value of b can be adjusted flexibly.

[0101] S7: Anti-counterfeiting test of fingerprint representation using descent gradient matrix and post-processing model.

[0102] First, the registered fingertip is tested. The fingertip is pressed onto the fingerprint acquisition ultrasonic transducer, and step S3 is executed to obtain the difference in the frequency domain acoustic field matrix generated after the fingertip is pressed. Then, the descent gradient matrix generated during offline training in step S4 is downloaded, and online fingerprint inversion is performed to obtain the imaging result, such as... Figure 11 As shown, its imaging range is 120mm × 120mm. To evaluate the imaging quality, step S2 is performed to obtain a fingerprint velocity model. Using the fingerprint velocity model as a benchmark, the root mean square error (RMSE) of the imaging result is 8.65 × 10⁻⁶. -3 This indicates that the imaging accuracy is high.

[0103] The imaging range was traced to a 36mm × 36mm area surrounding the fingerprint, and the DLCNN model built in step S5 was used for image inpainting. After inpainting, artifacts in the image were effectively removed. Subsequently, the MaskR-CNN model built in step S5 was used for fingerprint detail matching, successfully identifying 5 large-area details. Using the personal authentication confidence algorithm established in step S6, the authentication confidence of this registered fingertip example was calculated to be 0.94, thus passing personal authentication.

[0104] To test the anti-counterfeiting performance, an authentication attack was performed using someone else's fingertip; the testing steps were the same as for registering the fingertip. Within an imaging area of ​​120mm × 120mm, the RMSE value was 1.35 × 10⁻⁶. -2 After image inpainting using the DLCNN model and fingerprint detail matching using the Mask R-CNN model, the calculated authentication confidence is 0.30, which is insufficient for personal authentication.

[0105] The test results above demonstrate that, through offline training, the ultrasonic guided wave fingerprint imaging method proposed in this invention effectively learns the nonlinear mapping relationship between fingerprint morphological changes caused by fingertip pressure and acoustic field matrix changes, thereby accurately characterizing the ridge and valley distribution of the fingerprint. Furthermore, the image post-processing model possesses image restoration and fingerprint detail matching capabilities, thus enhancing authentication and anti-counterfeiting features.

[0106] S8: High-sensitivity fingerprint characterization test, resulting in a high-sensitivity fingerprint authentication method.

[0107] like Figure 12 As shown, when the fingertip contains moisture or dust, the quality of the fingerprint image collected by the optical sensor is significantly affected, exhibiting decreased contrast between ridges and valleys or discontinuous ridge and valley morphology. This indicates that traditional optical fingerprint scanning methods are easily affected by environmental pollution.

[0108] In contrast, the ultrasonic guided wave fingerprint imaging method proposed in this invention, after repeating the same operation of registering the fingertip test in step S7, shows that the retrieved fingerprint morphology has resistance to interference from moisture or dust, with an RMSE value of 1.76 × 10⁻⁶. -2 and 1.09×10 -2 The prediction accuracy is relatively high. After image restoration and fingerprint detail matching, the authentication confidence levels are 0.73 and 0.80, respectively, both passing personal authentication.

[0109] The test results under the above environmental pollution conditions show that the ultrasonic guided wave fingerprint imaging method proposed in this invention can effectively resist the impact of environmental pollution on imaging and authentication. While achieving zero space occupation under the transducer screen, it also ensures high-sensitivity imaging capability and authentication anti-counterfeiting performance.

[0110] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A highly sensitive and anti-counterfeiting ultrasonic guided wave fingerprint imaging and authentication method, characterized in that, The steps include: S1: Construct an ultrasonic transducer for fingerprint acquisition and place the ultrasonic transducer on an optical sensor. The ultrasonic transducer acquires the acoustic field signal of the fingerprint, and the optical sensor acquires the grayscale image of the fingerprint. The ultrasonic transducer includes several ultrasonic transducer elements arranged in a circular ring. Each ultrasonic transducer element includes a piezoelectric ceramic sheet group (1). A positive copper foil (2) and a negative copper foil (3) are respectively arranged on the upper and lower sides of the piezoelectric ceramic sheet group (1). A glass plate (4) is arranged on the lower side of the negative copper foil (3) and the upper side of the positive copper foil. Both the positive copper foil (2) and the negative copper foil (3) are connected to the ultrasonic instrument. An insulating layer (8) is laid between the positive copper foil (2) and the negative copper foil (3). A sound field boundary absorption layer (9) is arranged at the outermost edge between the two glass plates. The piezoelectric ceramic sheet group (1) is composed of at least one circular PZT-5H sheet. The positive electrode copper foil (2) and the negative electrode copper foil (3) both include circular endpoints with the same diameter as the PZT-5H sheet. Leads are provided on the circular endpoints. The ends of the leads of the negative electrode copper foil (3) are connected to the annular copper foil, which is grounded. The outermost part of the negative electrode copper foil (3) is directly connected to a square copper foil. The four corners of the square copper foil are divided into straight copper foil and right-angle copper foil by the break points. The straight copper foil is folded 90° outward, and the copper surface is flipped to face downward and on the same side as the copper surface of the positive electrode copper foil (2). The right-angle copper foil is reserved as a debugging contact point (10). The positive electrode copper foil (2) is divided into four parts, with the copper-clad surface of each part facing down; each part of the positive electrode copper foil is fixed with a polydimethylsiloxane film (7); the outermost periphery of the glass plate (4) is surrounded by modeling clay to serve as a sound field boundary absorption layer (9); A transparent polydimethylsiloxane film is laid between the surface glass of the optical sensor and the glass plate (4) of the ultrasonic transducer; S2: Preprocess the fingerprint grayscale image and calculate the phase velocity range to characterize the fingerprint based on the global matrix method. Establish a fingerprint velocity model based on the preprocessed fingerprint grayscale image. S3: Extract the first arrival wave time domain signal segment of the phase velocity interval corresponding to the pattern in the acoustic field signal of the fingerprint, extract the frequency domain energy flow amplitude to obtain the spectrum amplitude, and establish the frequency domain acoustic field matrix; S4: Build an offline training dataset for fingerprint imaging using the methods in steps S2 and S3, obtain the descent gradient matrix through offline training; build and train a post-processing model for image inpainting and detail matching. The supervised descent method used in the offline training is: In the formula, R is the descent gradient matrix obtained through offline training, Δs = s0 - s1 is the difference in the sound field matrix, where s0 and s1 represent the frequency domain sound field matrix before and after fingertip pressing, respectively; β is the regularization weight, Δm = m0 - m1 is the difference in the fingerprint velocity model, where m0 and m1 represent the fingerprint velocity model before and after fingertip pressing, respectively; the velocity sound field matrix s and the fingerprint velocity model m are both data matrices containing all examples; The post-processing model includes a DLCNN model for image restoration and a Mask R-CNN model for fingerprint matching. The DLCNN model adopts a small-scale, deep architecture, with all convolutional layers using a 3×3 dimension. The Mask R-CNN model includes a ResNet101 feature extraction network, a feature pyramid network, a region proposal network, a region of interest arrangement, and a fully convolutional network connected in sequence. The restored fingerprint grayscale image is obtained using the DLCNN model, and then converted into a binary form of the fingerprint image using an even-symmetric Gabor filter before being input into the Mask R-CNN model. The restored fingerprint grayscale image is then processed by the Mask R-CNN model to obtain the fingerprint matching result image. The training dataset for the Mask R-CNN model includes binarized grayscale images of fingerprints from multiple registered fingertips acquired by an optical sensor, with fingerprint markers attached. S5: Acquire the acoustic field signal of the fingerprint to be authenticated through an ultrasonic transducer, acquire the grayscale image of the fingerprint to be authenticated through an optical sensor, obtain the fingerprint velocity model and frequency domain acoustic field matrix of the test case using the methods in steps S2 and S3, obtain the fingerprint inversion image of the test case through online inversion of the descent gradient matrix; input the fingerprint inversion image into the trained post-processing model to obtain the fingerprint matching result. S6: Calculate personal authentication confidence using fingerprint matching results, and identify the fingerprint to be authenticated based on the personal authentication confidence; A wide-range marking method covering multiple fingerprint details was adopted, marking 8 wide-range fingerprint details; The method for calculating personal authentication confidence is as follows: the uniqueness of fingerprint details of the registered fingertip is divided into three levels. Level 1 fingerprint details have the highest uniqueness and are selected as key features representing personal information. Level 2 fingerprint details have the next highest uniqueness, while Level 3 fingerprint details have relatively weaker uniqueness. Therefore, the personal authentication confidence is: In the formula, ξ k The compensation factor is an empirical value related to the labeled fingerprint details, which varies with the number of fingerprint details k∈[0,8] identified by the Mask R-CNN model, and b is the weight of level 1 to 3.

2. The high-sensitivity and anti-counterfeiting ultrasonic guided wave fingerprint imaging and authentication method according to claim 1, characterized in that, The positive electrode copper foil (2) and the negative electrode copper foil (3) are respectively fixedly connected to the PZT-5H sheet by conductive silver paste (6), and the negative electrode copper foil (3) is fixed to the glass plate (4) on the lower side of the negative electrode copper foil (3) by epoxy resin adhesive (5).

3. The high-sensitivity and anti-counterfeiting ultrasonic guided wave fingerprint imaging and authentication method according to claim 2, characterized in that, The PZT-5H sheet is made by cutting and polishing the PZT-5H blank, and the positive electrode copper foil (2) and the negative electrode copper foil (3) are both obtained by laser cutting.

4. The high-sensitivity and anti-counterfeiting ultrasonic guided wave fingerprint imaging and authentication method according to any one of claims 1-3, characterized in that, The method for preprocessing the fingerprint grayscale image is as follows: fingerprint tracking and preliminary feature extraction are performed to obtain preliminary features of the fingerprint image; target regions in the preliminary features of the fingerprint image are extracted by threshold control; the extracted target regions are converted into low dynamic range images using a bilateral filtering algorithm; and connected regions in fingerprint details are eliminated by a spatial filtering algorithm with convolution kernels to obtain fingerprint ridge images.

5. The high-sensitivity and anti-counterfeiting ultrasonic guided wave fingerprint imaging and authentication method according to claim 4, characterized in that, The method for establishing the fingerprint velocity model is as follows: using the thickness of the fingertip epithelium, the height difference between the ridges and valleys, the epithelial density, Young's modulus, and Poisson's ratio as inputs, the phase velocity dispersion curves before and after the fingertip ridge is attached to the glass plate of the ultrasonic transducer are calculated using the global matrix method. At the same frequency, the A0 mode, which has a shorter wavelength and carries more acoustic field information, is selected. Compared with the glass structure of the glass plate, the material properties of the fingertip epithelium are closer to those of a fluid. Fluid-structure interaction theory shows that after the ridge is attached to the glass, the acoustic field energy flow is mainly concentrated in the adjacent mode below the phase velocity dispersion curve of the glass plate. At the center frequency of the ultrasonic transducer, the attachment of the ridge to the glass plate causes the phase velocity of the A0 mode to decrease from 2.91 mm / μs to 2.55 mm / μs, resulting in a phase velocity range of [2.55, 2.91] mm / μs. The fingerprint ridge image is mapped to the phase velocity range [2.55, 2.91] mm / μs to obtain the fingerprint velocity model.

6. The high-sensitivity and anti-counterfeiting ultrasonic guided wave fingerprint imaging and authentication method according to claim 5, characterized in that, The method for extracting the first arrival time domain signal segment of the phase velocity interval corresponding to the fingerprint acoustic field signal is as follows: the group velocity dispersion curve before and after the ridge is attached to the glass plate is calculated by using the conversion relationship between phase velocity and group velocity. At the center frequency of the ultrasonic transducer, there is an obvious mode separation phenomenon between the A0 mode and other modes. The time domain signal of the A0 mode is extracted from the time domain signal of the collected acoustic field signal by using the group velocity difference and window function. The method for extracting the frequency domain energy flow amplitude to obtain the spectral amplitude is as follows: the time domain signal segment of each ultrasonic transducer array element is transformed into a frequency domain signal by frequency domain transformation, the frequency domain signal is smoothed, the peak value of the frequency domain signal spectrum after smoothing is extracted to obtain the spectral amplitude, the extracted spectral amplitude is written into the corresponding point of the frequency domain sound field matrix, and the frequency domain sound field matrix is ​​mapped to the interval [1,2].

7. The high-sensitivity and anti-counterfeiting ultrasonic guided wave fingerprint imaging and authentication method according to claim 5, characterized in that, The imaging result M of the test case is obtained by performing online inversion on the descent gradient matrix R. test =m0+Δs test ·R;Δs test This is the frequency domain sound field matrix corresponding to the test case.

8. The high-sensitivity and anti-counterfeiting ultrasonic guided wave fingerprint imaging and authentication method according to claim 7, characterized in that, The method for fingerprint tracking and preliminary feature extraction is as follows: In the fingerprint grayscale image, track the point with the largest grayscale value, take the point with the largest grayscale value as the center, extract the grayscale image within a range of 36mm×36mm around it, invert the grayscale image, and extract the range with grayscale values ​​below 110; The threshold control method is to discretize the inverted grayscale image with grid spacing, and retain the part with an average grayscale value below 80 within the grid. The smoothing process uses a Gaussian weighted moving average filtering algorithm; The method for establishing the offline training dataset is as follows: collect sound field signals and fingerprint grayscale images of the fingertip epithelium under different dry conditions, at different locations, and with different pressures. The fingerprint velocity model and sound field signal obtained for each fingertip press are combined into a calculation example through the methods in steps S2 and S3.

9. The high-sensitivity and anti-counterfeiting ultrasonic guided wave fingerprint imaging and authentication method according to claim 8, characterized in that, The DLCNN model has a network depth of 20 layers. The training dataset of the DLCNN model includes grayscale fingerprint images for artifact removal and grayscale images of natural landscape images for image denoising. Each example in the artifact removal part includes an image of fingerprint inversion with artifacts and a corresponding clean fingerprint image. Each example in the image denoising part includes a natural landscape image with Gaussian noise and a corresponding clean image.

Images