A fingerprint identification method based on white-box and black-box joint evaluation

Abstract

The application discloses a kind of based on white box-black box joint evaluation fingerprint identification method, comprising: according to the multiple real acquisition conditions of pre-set, obtain multi-condition semiconductor fingerprint dataset, wherein the multiple real acquisition conditions include: different finger humidity conditions and different pressing pressure conditions;According to the multi-condition semiconductor fingerprint dataset, introduce uncertainty evaluation model, calculate the performance degradation rate of acquisition module, feature extraction module and matching module in fingerprint model under environmental disturbance, obtain the uncertainty index corresponding to each module;According to the uncertainty index, the white box evaluation result of acquisition module, feature extraction module and matching module is associated with system level black box performance index Analysis, obtain fingerprint identification result.The method constructs multidimensional evaluation system, can more comprehensively reflect the comprehensive performance of algorithm.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to a fingerprint recognition method based on white-box and black-box joint evaluation. Background Technology

[0002] With the rapid development of the smart door and smart lock industry, semiconductor fingerprint recognition has become one of the most crucial identity authentication methods for electronic smart locks. The performance of fingerprint recognition algorithms directly affects product security, user experience, and market application reliability. Current fingerprint algorithm performance evaluation typically employs end-to-end "black-box testing," focusing only on the relationship between system input (fingerprint image) and output (matching score / authentication result), using FAR (False Acceptance Rate) and FRR (False Rejection Rate) as primary indicators to evaluate overall system performance. The internationally used FVC (Fingerprint Verification Competition) testing system also employs these black-box metrics to evaluate algorithms.

[0003] However, black-box evaluation has a fundamental limitation: when the overall system performance is poor, it is impossible to determine whether the root cause lies in fingerprint image acquisition, feature extraction, or matching, leading to a lack of targeted direction for algorithm optimization. In recent years, the academic community has also proposed the concept of white-box evaluation, advocating for the independent evaluation of the performance of each sub-module of the system. However, existing white-box research mostly focuses on a single module (such as evaluating only the collector or only the feature extractor), lacking a complete framework for unified white-box evaluation of the acquisition, feature extraction, and matching modules and correlation analysis with black-box results.

[0004] Existing fingerprint algorithm testing methods have the following drawbacks: First, existing methods are all end-to-end black-box testing or only module-level white-box testing. Black-box evaluation cannot pinpoint the problem area, and white-box evaluation is performed independently for each module and is not related to the black-box results. When the algorithm misidentifies or rejects an item, developers cannot determine whether it is due to a decline in image acquisition quality, incorrect feature point extraction, or insufficient robustness of the matching algorithm, and therefore cannot accurately pinpoint the direction for optimization.

[0005] Second, existing methods lack quantitative analysis of the uncertainties of each module, and there is no unified method for calculating uncertainty that can be compared across modules, which makes it impossible to directly compare the stability of different modules.

[0006] Third, the angle of finger placement varies during door lock use, users may have dry or wet hands, and the pressing pressure varies greatly. Existing general evaluation methods do not have specific test conditions designed for them, and cannot evaluate the differences in sensitivity of the collector, feature extractor, and matcher to working conditions such as dry hands, wet hands, and pressure changes, making it difficult to guide module selection and combination optimization. Summary of the Invention

[0007] To address the aforementioned technical issues, this invention proposes a fingerprint recognition method based on white-box and black-box joint evaluation. This method constructs a multi-dimensional evaluation system that can more comprehensively reflect the overall performance of the algorithm.

[0008] To achieve the above objectives, the present invention provides a fingerprint recognition method based on white-box and black-box joint evaluation, comprising: Based on a variety of preset real-world acquisition conditions, a multi-condition semiconductor fingerprint dataset is obtained, wherein the various real-world acquisition conditions include: different finger humidity conditions and different pressing pressure conditions. Based on the multi-condition semiconductor fingerprint dataset, an uncertainty assessment model is introduced to calculate the performance degradation rate of the acquisition module, feature extraction module, and matching module in the fingerprint model under environmental disturbances, and obtain the uncertainty index corresponding to each module. Based on the uncertainty index, the white-box evaluation results of the acquisition module, feature extraction module and matching module are correlated with the system-level black-box performance index to obtain the fingerprint recognition result.

[0009] Optionally, based on various preset real-world acquisition conditions, a multi-condition semiconductor fingerprint dataset can be obtained, including: Based on publicly available datasets and existing standard image libraries, fingerprint images of the same batch of fingers were collected under various real-world acquisition conditions. Based on the fingerprint images acquired under the various real acquisition conditions, a multi-condition semiconductor fingerprint dataset is formed, which includes reference samples and perturbation samples. The reference samples are fingerprint images obtained under normal acquisition conditions, and the perturbation samples are homologous fingerprint images obtained under dry, humid, low-pressure, or high-pressure conditions.

[0010] Optionally, the different finger humidity conditions include a dry state, a normal state, and a moist state, and the different pressing pressure conditions include a low pressure state, a normal pressure state, and a high pressure state.

[0011] Optionally, based on the multi-condition semiconductor fingerprint dataset, an uncertainty assessment model is introduced to calculate the performance degradation rate of the acquisition module, feature extraction module, and matching module in the fingerprint model under environmental disturbances, and to obtain the uncertainty index corresponding to each module, including: Based on the reference sample and the perturbation sample, the module scores of the acquisition module, feature extraction module and matching module under the perturbation condition are calculated respectively, and the module scores are normalized to a preset interval to obtain the degradation rate sequence. Based on the degradation rate sequence, calculate the mean and standard uncertainty of each reference sample, and summarize all reference samples to obtain the overall standard uncertainty, and obtain the uncertainty index corresponding to each module.

[0012] Optionally, the module score for the data acquisition module under disturbance conditions includes: Based on the reference sample and the perturbation sample, the fingerprint images acquired under different conditions are quantitatively analyzed using image quality evaluation indicators to obtain the quantitative analysis results. Based on the quantitative analysis results, fingerprint images collected under normal conditions are used as reference samples, and homologous fingerprint images collected under different perturbation conditions are used as perturbation samples. The fluctuation degree of the quality score during the perturbation process is calculated to obtain the module score of the acquisition module under perturbation conditions.

[0013] Optionally, the module score for the feature extraction module under perturbation conditions includes: Based on the reference sample and the perturbation sample, a standard minutiae reference database is established according to the reference fingerprint image with manually annotated real minutiae feature positions and directions. Homologous fingerprint images under different environmental conditions are input into the feature extraction algorithm to be evaluated to obtain an automatically extracted set of minutiae. The automatically extracted set of minutiae is then paired with a standard minutiae reference database using a neighborhood matching strategy. Based on the pairing results, the minutiae detection capability index, root mean square error of position, and orientation error are calculated, and the performance parameter degradation rate under different environmental conditions is statistically analyzed to obtain the uncertainty index of the feature extraction module.

[0014] Optionally, calculating the module score of the matching module under perturbation conditions includes: Based on the reference samples and perturbation samples, and combined with the standard minutiae reference database, a minutiae perturbation model conforming to real application scenarios is constructed. The minutiae perturbation model includes a local occlusion perturbation model and a global rotation perturbation model. The local occlusion perturbation model is used to simulate different degrees of local information loss, and the global rotation perturbation model is used to simulate different degrees of attitude shift. A perturbation template is generated based on the minutiae perturbation model, and the reference template and the perturbation template are input into the matching algorithm to be evaluated to obtain the corresponding matching score; Statistical analysis is performed based on the changes in the matching scores to obtain the uncertainty index of the matching module.

[0015] Optionally, based on the uncertainty index, the white-box evaluation results of the acquisition module, feature extraction module, and matching module are correlated with the system-level black-box performance index to obtain fingerprint recognition results, including: Based on the uncertainty index, the sensitivity of each module to environmental disturbances is determined. Based on the sensitivity of each module to environmental disturbances, and combined with system-level black-box performance indicators, the target module causing the system performance degradation is located, so as to achieve a comprehensive evaluation of the fingerprint recognition system.

[0016] Compared with the prior art, the present invention has the following advantages and technical effects: 1. Existing evaluation methods often employ artificially synthesized noise, motion blur, or idealized rotation, which significantly differ from real-world user scenarios involving dry or wet hands, varying pressure applied, and finger placement angles. The evaluation results of this invention more accurately reflect the algorithm's performance in actual products. 2. The uncertainty assessment model proposed in this invention, through normalized performance scores and standard deviation summaries, yields a dimensionless and range-stable overall uncertainty. This index is applicable to different types of evaluation outputs, such as acquisition quality, feature detection accuracy, and matching scores, thus achieving a unified quantitative comparison of the robustness of heterogeneous modules such as collectors, feature extractors, and matchers. This has clear practical value in technology selection, bidding evaluation, and algorithm iteration. 3. Traditional evaluations often focus only on a single index (such as matching score or EER), easily masking certain algorithmic shortcomings. This invention constructs a multi-dimensional evaluation system that can more comprehensively reflect the algorithm's overall performance. Attached Figure Description

[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a fingerprint recognition method based on white-box-black-box joint evaluation according to an embodiment of the present invention. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0020] This embodiment proposes a fingerprint recognition method based on white-box and black-box joint evaluation, such as... Figure 1 As shown, the specific steps include: Based on a variety of preset real-world acquisition conditions, a multi-condition semiconductor fingerprint dataset is obtained, wherein the various real-world acquisition conditions include: different finger humidity conditions and different pressing pressure conditions. Based on the multi-condition semiconductor fingerprint dataset, an uncertainty assessment model is introduced to calculate the performance degradation rate of the acquisition module, feature extraction module, and matching module in the fingerprint model under environmental disturbances, and obtain the uncertainty index corresponding to each module. Based on the uncertainty index, the white-box evaluation results of the acquisition module, feature extraction module and matching module are correlated with the system-level black-box performance index to obtain the fingerprint recognition result.

[0021] Specifically, 1. This invention constructs a standard image library and interference image libraries such as dry and wet, high pressure, and low pressure, and uses the uncertainty method to evaluate the three core modules of the fingerprint recognition system respectively.

[0022] 2. White-box evaluation schemes closely resembling real-world usage scenarios were designed for the three core modules: The data acquisition module combines statistical hypothesis testing (t-test) and uncertainty models to quantify the collector's sensitivity to environmental changes; the feature extraction module transfers stylistic features from real-world collected images (wet fingers, dry fingers, high pressure, etc.) to standard fingerprint images with manually annotated details, generating realistic perturbation test samples. This method is closer to actual door lock usage scenarios than traditional artificial noise or motion blur. The matching module addresses common issues in door lock use, such as missing details and posture deviations. It designs a local occlusion perturbation model to simulate ridge adhesion and block detail loss caused by wet fingers, and a global rotation perturbation model to simulate different finger placement angles. Perturbation intensity, such as occlusion area and rotation angle, is set according to levels to evaluate the robustness of the matching algorithm.

[0023] 3. Unlike the independent end-to-end black-box testing or single-module white-box testing in existing technologies, this invention is the first to construct a unified framework that correlates the white-box evaluation results of the acquisition module, feature extraction module and matching module with system-level black-box performance indicators, thereby achieving comprehensive diagnosis and optimization guidance for fingerprint recognition systems.

[0024] Furthermore, based on various preset real-world acquisition conditions, a multi-condition semiconductor fingerprint dataset is obtained, including: Based on publicly available datasets and existing standard image libraries, fingerprint images of the same batch of fingers were collected under various real-world acquisition conditions. Based on the fingerprint images acquired under the various real acquisition conditions, a multi-condition semiconductor fingerprint dataset is formed, which includes reference samples and perturbation samples. The reference samples are fingerprint images obtained under normal acquisition conditions, and the perturbation samples are homologous fingerprint images obtained under dry, humid, low-pressure, or high-pressure conditions.

[0025] Furthermore, the different finger humidity conditions include a dry state, a normal state, and a moist state, and the different pressing pressure conditions include a low pressure state, a normal pressure state, and a high pressure state.

[0026] Specifically, step 1: Construct a multi-condition semiconductor acquisition dataset. Based on publicly available datasets and existing standard image libraries, fingerprint images of the same batch of fingers are acquired under various real-world acquisition conditions. These acquisition conditions include: different finger humidity levels (dry, normal, wet) and different pressing pressures (low, normal, high). An internal fingerprint dataset acquired under different interference conditions (humidity, pressure, and light) is utilized. To control finger humidity, a skin humidity meter can be used to record the percentage of skin surface humidity. The recorded humidity categories include the subject's natural skin humidity (normal state), humidity after wiping the fingertips with a dry tissue (dry state), and humidity after applying a small amount of moisturizing lotion to the fingertips. Finally, this embodiment uses a tactile grip and pressure sensing system to measure the applied pressure values ​​under normal pressure, low pressure, and high pressure conditions. In the normal pressure test, the subject needs to place the fingerprint on the reader surface without consciously controlling the pressure. In the low pressure test, the subject only needs to place the finger flat on the imaging surface and avoid applying downward pressure as much as possible. The high pressure test requires the subject to place the finger on the imaging surface while applying maximum downward pressure.

[0027] Furthermore, based on the multi-condition semiconductor fingerprint dataset, an uncertainty assessment model is introduced to calculate the performance degradation rate of the acquisition module, feature extraction module, and matching module in the fingerprint model under environmental disturbances, and to obtain the uncertainty indices corresponding to each module, including: Based on the reference sample and the perturbation sample, the module scores of the acquisition module, feature extraction module and matching module under the perturbation condition are calculated respectively, and the module scores are normalized to a preset interval to obtain the degradation rate sequence. Based on the degradation rate sequence, calculate the mean and standard uncertainty of each reference sample, and summarize all reference samples to obtain the overall standard uncertainty, and obtain the uncertainty index corresponding to each module.

[0028] Specifically, step 2: In order to quantitatively analyze the robustness and stability of each module of the fingerprint recognition system in real application scenarios, this embodiment introduces a new uncertainty assessment method. The proposed method takes into account the performance deviation caused by disturbance conditions, thereby more comprehensively reflecting the sensitivity of the algorithm to environmental changes.

[0029] First, a reference fingerprint sample set is constructed under standard acquisition conditions. Let there be M reference fingerprint samples, and their corresponding feature representations be denoted as... (k=1,2,...,M). Reference samples were collected under normal illumination, natural humidity, and moderate pressing pressure conditions to ensure they represent the system's optimal operating state. Subsequently, for each reference sample, N corresponding samples were obtained under different perturbation conditions, including typical operating conditions such as dry fingers, wet fingers, low pressure, and high pressure. The feature representation extracted under perturbation conditions is denoted as... (n=1,2,...,N), calculate and The module scores were determined by the differences in the dimensions of different evaluation indicators. To eliminate the discrepancies in the dimensions of different evaluation indicators, the scores were normalized to the [0,1] interval. .

[0030] For different evaluation modules, performance metrics corresponding to reference samples and perturbation samples are calculated. For the image acquisition module, existing image quality evaluation metrics can be used; for the feature extraction module, metrics such as position error and orientation error can be used; for the matching module, performance parameters such as matching score, FNMR, or EER are used; and for the system as a whole, the average matching score of the real matching pairs is used.

[0031] Based on the above definition, a set of degradation rate sequences can be obtained for the same reference sample under different perturbation conditions. , ,..., For each reference sample, calculate the mean of the N ratings: ; The standard uncertainty of the reference sample was calculated as follows: ; Repeating the above steps for all reference samples yields the population standard uncertainty defined as: ; The resulting expanded uncertainty not only reflects the performance stability of the algorithm under complex environmental conditions, but also provides a unified quantitative basis for objective comparison between different fingerprint recognition modules, which is more in line with the needs of robustness evaluation of fingerprint recognition systems in actual door lock applications.

[0032] Furthermore, the module score for the data acquisition module under disturbance conditions includes: Based on the reference sample and the perturbation sample, the fingerprint images acquired under different conditions are quantitatively analyzed using image quality evaluation indicators to obtain the quantitative analysis results. Based on the quantitative analysis results, fingerprint images collected under normal conditions are used as reference samples, and homologous fingerprint images collected under different perturbation conditions are used as perturbation samples. The fluctuation degree of the quality score during the perturbation process is calculated to obtain the module score of the acquisition module under perturbation conditions.

[0033] Specifically, step 3 involves evaluating the impact of the fingerprint scanner itself on fingerprint image quality. To this end, using the interference dataset constructed above, existing image quality evaluation indicators can be employed to quantitatively analyze fingerprint images acquired under different conditions. A t-test with a significance level of α=0.05 is used to verify whether fingerprint dryness / wetness and pressure disturbance conditions lead to a statistically significant decrease in image quality. Based on this, an uncertainty assessment method can be used, using fingerprint images acquired under normal conditions as reference samples and the same fingerprint images acquired under different disturbance conditions as disturbance samples. By calculating the fluctuation of the quality score during the disturbance process, the sensitivity of the acquisition device to environmental changes can be quantitatively reflected.

[0034] Furthermore, the module score for the feature extraction module under perturbation conditions includes: Based on the reference sample and the perturbation sample, a standard minutiae reference database is established according to the reference fingerprint image with manually annotated real minutiae feature positions and directions. Homologous fingerprint images under different environmental conditions are input into the feature extraction algorithm to be evaluated to obtain an automatically extracted set of minutiae. The automatically extracted set of minutiae is then paired with a standard minutiae reference database using a neighborhood matching strategy. Based on the pairing results, the minutiae detection capability index, root mean square error of position, and orientation error are calculated, and the performance parameter degradation rate under different environmental conditions is statistically analyzed to obtain the uncertainty index of the feature extraction module.

[0035] Specifically, step 4: To quantitatively analyze the stability, accuracy, and anti-interference capability of the feature extraction module in a fingerprint recognition system under complex application environments, this invention uses existing research methods. Specifically, in this embodiment, a perturbation is first applied to the fingerprint image with manually marked real minutiae positions and orientations, and then the fingerprint image is input into the minutiae extractor. Building upon existing perturbation techniques such as motion blur and random noise, this invention uses a realistic perturbation method to simulate the impact of wet fingers, dry fingers, and different pressing pressure environments on the fingerprint image during actual use. Furthermore, it utilizes minutiae detection capability, positioning accuracy, and uncertainty analysis to achieve a comprehensive evaluation of the feature extraction algorithm's performance.

[0036] First, a standard minutiae reference database is established. For reference fingerprint images, professional annotators complete the annotation of real minutiae based on the position and direction of ridge termination points and ridge bifurcation points, forming a standard minutiae set. This set is considered as a true representation of features under ideal conditions and is used as a benchmark reference in subsequent evaluation processes.

[0037] For the k-th reference fingerprint image, its set of true minutiae is represented as: ; in, This represents the actual number of detail points; For the coordinates of the detail points; For the direction angle of the detail point.

[0038] Subsequently, the homologous fingerprint images obtained in step 1 under different environmental conditions are input into the feature extraction algorithm to be evaluated to obtain an automatically extracted set of minutiae, as shown in the following formula: ; in, To automatically extract the number of detail points.

[0039] To establish the correspondence between automatically extracted minutiae and manually annotated minutiae, a neighborhood matching strategy is employed. A successful match is determined when the Euclidean distance between the extracted minutiae and the actual minutiae is less than a preset threshold δ. ≤δ; If multiple extracted minutiae simultaneously satisfy the distance constraint, the minutiae with the smallest distance and the smallest directional difference is selected as the final matching result. Real minutiae that cannot be matched are defined as missed minutiae; extracted minutiae that cannot be matched with any real minutiae are defined as pseudo minutiae.

[0040] In terms of evaluating minutiae detection capabilities, the number of successfully matched minutiae, the number of missed minutiae, and the number of false minutiae are statistically analyzed, and the overall detection performance is evaluated using the GI index, as shown in the following formula: ; in: The quality weight of the l-th image patch is... To accurately detect the number of detail points, To accurately detect the number of detail points, To avoid missing details, This represents the true minutiae count. A higher GI value indicates a closer approximation between the extracted minutiae and the actual set of minutiae, resulting in better detection performance. For evaluating minutiae localization accuracy, the root mean square error of the position and the minutiae orientation error are calculated for all successfully matched minutiae. ; ; This metric measures the positional and directional deviation between automatically extracted detail points and actual detail points; the smaller the value, the higher the positioning accuracy.

[0041] Further analysis of the impact of environmental factors on feature extraction performance was conducted. The GI, positional error, and orientation error parameters were statistically analyzed under normal conditions, dry hand conditions, wet hand conditions, low pressure conditions, and high pressure conditions. The degradation rates were as follows: , , ; For positive evaluation metrics such as GI, the degree of degradation is represented by the percentage decrease in performance; for negative evaluation metrics such as position error and orientation error, the degree of degradation is represented by the percentage increase in error. The final uncertainty score for this module can be obtained using the method in step 2. ; Through the above evaluation process, multi-dimensional performance indicators such as minutiae detection capability, positioning accuracy, orientation estimation accuracy, and environmental robustness can be obtained simultaneously, and the sensitivity of the feature extraction module to changes in real-world door lock usage scenarios can be further quantified. Compared with existing evaluation methods based on artificial noise and motion blur, this invention uses real-world application environment disturbances to establish an evaluation system, enabling the evaluation results to more accurately reflect the performance of the smart door lock fingerprint algorithm in practical engineering applications, and providing a reliable data foundation for subsequent evaluation of the matching module and comprehensive evaluation of the entire system.

[0042] Furthermore, calculating the module score of the matching module under perturbation conditions includes: Based on the reference samples and perturbation samples, and combined with the standard minutiae reference database, a minutiae perturbation model conforming to real application scenarios is constructed. The minutiae perturbation model includes a local occlusion perturbation model and a global rotation perturbation model. The local occlusion perturbation model is used to simulate different degrees of local information loss, and the global rotation perturbation model is used to simulate different degrees of attitude shift. A perturbation template is generated based on the minutiae perturbation model, and the reference template and the perturbation template are input into the matching algorithm to be evaluated to obtain the corresponding matching score; Statistical analysis is performed based on the changes in the matching scores to obtain the uncertainty index of the matching module.

[0043] Specifically, step 5: To further analyze the sensitivity of the matching module in the fingerprint recognition system to feature perturbations and its stability in complex application environments, this invention proposes a white-box evaluation method for the matching module based on minutiae feature space perturbations. This method uses the standard minutiae template obtained in step 4 as a reference feature set. By constructing a minutiae perturbation model that conforms to real application scenarios, it simulates the impact of feature extraction errors, fingerprint image degradation, and actual user behavior on the matching process, thereby achieving a quantitative evaluation of the robustness of the matching algorithm.

[0044] First, minutiae templates are extracted from a standard fingerprint database containing manually annotated real minutiae location and orientation information to form a reference minutiae set: ; in, =( ) indicates the first The location coordinates and orientation information of each detailed point.

[0045] Considering that factors such as wet fingers, dry fingers, high-pressure pressing, and different placement angles can lead to phenomena such as missing details, partial obstruction, and overall rotation during actual door lock use, this invention combines two types of perturbation models that conform to real application scenarios.

[0046] The first type is the local occlusion perturbation model. This model is used to simulate situations such as ridge adhesion, local image distortion, and feature extraction failure caused by wet fingers. Occlusion regions with gradually increasing areas are randomly generated in the minutiae space, and all minutiae within these regions are deleted to form the perturbation template. The area of ​​the occluded region is set in multiple increments, such as 32×32 pixels, 64×64 pixels, 128×128 pixels, 256×256 pixels, etc., to simulate different degrees of local information loss.

[0047] The second type is the global rotational perturbation model. This model is used to simulate the posture deviation caused by the user placing their finger and the impact of orientation estimation errors during feature extraction on the matching results.

[0048] For any minutiae in the original set of minutiae ( The updated direction angle is shown below: ; ; The rotation angles were set in multiple levels: 5°, 10°, 15°, and 20°, simulating slight, moderate, and severe attitude deviations, respectively. A reference template was then used. and By inputting the matching algorithm to be evaluated, the corresponding matching score can be obtained.

[0049] Furthermore, the uncertainty evaluation model established in step 2 is used to conduct statistical analysis on the changes in matching scores under different disturbance conditions, and the uncertainty index of the matching module is obtained, which can help understand the sensitivity and stability level of the detail point disturbance.

[0050] Furthermore, based on the aforementioned uncertainty index, the white-box evaluation results of the acquisition module, feature extraction module, and matching module are correlated with the system-level black-box performance index to obtain fingerprint recognition results, including: Based on the uncertainty index, the sensitivity of each module to environmental disturbances is determined. Based on the sensitivity of each module to environmental disturbances, and combined with system-level black-box performance indicators, the target module causing the system performance degradation is located, so as to achieve a comprehensive evaluation of the fingerprint recognition system.

[0051] Specifically, step 6: To evaluate the overall performance of the fingerprint recognition system in a real-world environment, this invention constructs an end-to-end black-box evaluation process. Using the multi-condition fingerprint database established in step 1, fingerprint images obtained under normal acquisition conditions are used as registration samples, while homologous fingerprint images obtained under dry, humid, high-pressure, and low-pressure conditions are used as verification samples. The verification samples sequentially pass through the image acquisition module, image preprocessing module, feature extraction module, and matching module to obtain the final matching result.

[0052] The false acceptance rate (FAR), false rejection rate (FRR), and equal error rate (EER) are calculated based on the matching results, and the recognition performance under different environmental conditions is also statistically analyzed. This evaluation process does not require access to the algorithm's internal parameters; it only utilizes the input-output relationship of the entire system to achieve a comprehensive evaluation of the system's performance, thus it is a black-box evaluation process.

[0053] Next, a white-box versus black-box correlation analysis can be performed. For example, if the system's end-to-end FNMR significantly increases under wet finger conditions, the white-box evaluation results can be reviewed. For instance, the acquisition results (Uacquisition) show that the collector is not sensitive to humidity, and the quality score does not decrease significantly; the feature results (Ufeature) show that the feature extractor's GI decreases significantly under the "wet finger" style, missing a large number of details; and the matching results (Umatching) show that the matcher has some robustness to local occlusion. Therefore, the performance bottleneck lies in the insufficient processing capability of the feature extraction module for wet finger images. This provides a precise direction for algorithm optimization.

[0054] The overall advantage of the system may stem from the superior performance of its acquisition or feature extraction modules, rather than from matching modules. This is also highly valuable for system integrators, who can select the optimal modules from different suppliers to combine based on specific application scenarios, such as whether they prioritize resistance to rotation or interference from dry and wet environments, rather than blindly choosing the single system that is optimal overall. Through steps 1-6 above, this invention forms a complete, multi-level, three-dimensional evaluation system, ranging from micro-modules to macro-systems, and from ideal conditions to real disturbances.

[0055] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A fingerprint recognition method based on white-box and black-box joint evaluation, characterized in that, include: Based on a variety of preset real-world acquisition conditions, a multi-condition semiconductor fingerprint dataset is obtained, wherein the various real-world acquisition conditions include: different finger humidity conditions and different pressing pressure conditions. Based on the multi-condition semiconductor fingerprint dataset, an uncertainty assessment model is introduced to calculate the performance degradation rate of the acquisition module, feature extraction module, and matching module in the fingerprint model under environmental disturbances, and obtain the uncertainty index corresponding to each module. Based on the uncertainty index, the white-box evaluation results of the acquisition module, feature extraction module and matching module are correlated with the system-level black-box performance index to obtain the fingerprint recognition result.

2. The fingerprint recognition method based on white-box and black-box joint evaluation according to claim 1, characterized in that, Based on various preset real-world acquisition conditions, a multi-condition semiconductor fingerprint dataset is obtained, including: Based on publicly available datasets and existing standard image libraries, fingerprint images of the same batch of fingers were collected under various real-world acquisition conditions. Based on the fingerprint images acquired under the various real acquisition conditions, a multi-condition semiconductor fingerprint dataset is formed, which includes reference samples and perturbation samples. The reference samples are fingerprint images obtained under normal acquisition conditions, and the perturbation samples are homologous fingerprint images obtained under dry, humid, low-pressure, or high-pressure conditions.

3. The fingerprint recognition method based on white-box and black-box joint evaluation according to claim 1, characterized in that, The different finger humidity conditions include dry state, normal state and wet state, and the different pressing pressure conditions include low pressure state, normal pressure and high pressure state.

4. The fingerprint recognition method based on white-box and black-box joint evaluation according to claim 2, characterized in that, Based on the multi-condition semiconductor fingerprint dataset, an uncertainty assessment model is introduced to calculate the performance degradation rate of the acquisition module, feature extraction module, and matching module in the fingerprint model under environmental disturbances, and to obtain the uncertainty indices corresponding to each module, including: Based on the reference sample and the perturbation sample, the module scores of the acquisition module, feature extraction module and matching module under the perturbation condition are calculated respectively, and the module scores are normalized to a preset interval to obtain the degradation rate sequence. Based on the degradation rate sequence, calculate the mean and standard uncertainty of each reference sample, and summarize all reference samples to obtain the overall standard uncertainty, and obtain the uncertainty index corresponding to each module.

5. The fingerprint recognition method based on white-box and black-box joint evaluation according to claim 4, characterized in that, The module score for the data acquisition module under disturbance conditions includes: Based on the reference sample and the perturbation sample, the fingerprint images acquired under different conditions are quantitatively analyzed using image quality evaluation indicators to obtain the quantitative analysis results. Based on the quantitative analysis results, fingerprint images collected under normal conditions are used as reference samples, and homologous fingerprint images collected under different perturbation conditions are used as perturbation samples. The fluctuation degree of the quality score during the perturbation process is calculated to obtain the module score of the acquisition module under perturbation conditions.

6. The fingerprint recognition method based on white-box and black-box joint evaluation according to claim 4, characterized in that, The module score for the feature extraction module under perturbation conditions includes: Based on the reference sample and the perturbation sample, a standard minutiae reference database is established according to the reference fingerprint image with manually annotated real minutiae feature positions and directions. Homologous fingerprint images under different environmental conditions are input into the feature extraction algorithm to be evaluated to obtain an automatically extracted set of minutiae. The automatically extracted set of minutiae is then paired with a standard minutiae reference database using a neighborhood matching strategy. Based on the pairing results, the minutiae detection capability index, root mean square error of position, and orientation error are calculated, and the performance parameter degradation rate under different environmental conditions is statistically analyzed to obtain the uncertainty index of the feature extraction module.

7. The fingerprint recognition method based on white-box and black-box joint evaluation according to claim 6, characterized in that, The calculation of the matching module's module score under perturbation conditions includes: Based on the reference samples and perturbation samples, and combined with the standard minutiae reference database, a minutiae perturbation model conforming to real application scenarios is constructed. The minutiae perturbation model includes a local occlusion perturbation model and a global rotation perturbation model. The local occlusion perturbation model is used to simulate different degrees of local information loss, and the global rotation perturbation model is used to simulate different degrees of attitude shift. A perturbation template is generated based on the minutiae perturbation model, and the reference template and the perturbation template are input into the matching algorithm to be evaluated to obtain the corresponding matching score; Statistical analysis is performed based on the changes in the matching scores to obtain the uncertainty index of the matching module.

8. The fingerprint recognition method based on white-box and black-box joint evaluation according to claim 1, characterized in that, Based on the uncertainty index, the white-box evaluation results of the acquisition module, feature extraction module, and matching module are correlated with the system-level black-box performance index to obtain fingerprint recognition results, including: Based on the uncertainty index, the sensitivity of each module to environmental disturbances is determined. Based on the sensitivity of each module to environmental disturbances, and combined with system-level black-box performance indicators, the target module causing the system performance degradation is located, so as to achieve a comprehensive evaluation of the fingerprint recognition system.

Images