A method and system for fast fingerprint retrieval based on image pre-screening

Abstract

The application relates to the technical field of image processing, in particular to a fingerprint rapid retrieval method and system based on image pre-screening, which comprises the following steps: obtaining a user clock-in request, obtaining a current fingerprint image, performing multi-dimensional quality evaluation on the current fingerprint image to obtain a quality feature vector, judging the quality feature vector according to a dynamic threshold corresponding to a fingerprint quality grade, performing residual image detection and prosthesis detection if all indexes meet the threshold, and entering a matching stage in response to the clock-in request; obtaining the quality feature vector, obtaining a quality score of the current image through a quality score regression model, selecting a corresponding matching strategy for fingerprint retrieval according to the quality score, recording attendance information if the matching is successful, and prompting the user to change fingers or use a backup verification mode if the matching fails. The application has the effects of improving the recognition efficiency and user experience for the fingerprint light crowd without increasing the hardware cost.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method and system for rapid fingerprint retrieval based on image pre-screening. Background Technology

[0002] Fingerprint attendance machines are widely used in enterprises and institutions for attendance management. They utilize the uniqueness and stability of human fingerprints for identity authentication. Compared to traditional methods such as card swiping and passwords, fingerprint attendance eliminates the need for physical media and is difficult to forge, greatly improving the convenience and security of attendance management. With the development of IoT technology, modern fingerprint attendance machines have integrated optical or semiconductor fingerprint sensors, embedded processors, and network communication modules. They can independently complete functions such as fingerprint collection, comparison, recording, and storage, and support data export and cloud synchronization, becoming an important component of smart offices.

[0003] The working principle of existing fingerprint attendance machines typically includes the following steps: the user presses their finger onto the fingerprint sensor, which captures the fingerprint image; the embedded system preprocesses the raw image, including image enhancement, binarization, and thinning operations, extracting fingerprint details (such as ridge endpoints and bifurcation points) as features; then, the extracted features are compared 1:N with a fingerprint template library pre-stored in the device, and attendance information is recorded if a match is found. Image pre-screening is a preliminary step in the preprocessing, designed to quickly remove low-quality images that are blurry, incomplete, improperly wetted or dry, or contain ghosting, avoiding unnecessary computation that consumes valuable processor resources, thereby improving overall recognition speed and accuracy.

[0004] However, in practical applications, the fingerprints of blue-collar workers, the elderly, and those engaged in long-term manual labor often exhibit characteristics such as severe wear, shallow lines, and dry, peeling skin. Traditional image pre-screening algorithms use fixed quality thresholds (such as clarity and coverage limits). When these users' fingerprint images fail to meet the preset quality score due to shallow lines, they are frequently judged as unqualified and require re-pressing, resulting in multiple failed attempts and a poor user experience. Furthermore, a uniform threshold can cause the pre-screening process to prematurely filter out a large number of low-quality but actually identifiable fingerprint images, increasing the burden on subsequent matching stages and even causing missed detections. Summary of the Invention

[0005] To improve recognition efficiency and user experience for people with faint fingerprints without increasing hardware costs, this application provides a fingerprint fast retrieval method and system based on image pre-screening.

[0006] The above-mentioned objective of this application is achieved through the following technical solution: A fast fingerprint retrieval method based on image pre-screening, the method comprising: Obtain a user registration request, obtain an original fingerprint image, and extract fingerprint feature data from the original fingerprint image; The fingerprint quality level is identified based on the fingerprint feature data, a corresponding clarity level is set based on the fingerprint quality level, and the user registration request is responded to. The system obtains a user's check-in request, acquires the current fingerprint image, performs a multi-dimensional quality assessment on the current fingerprint image to obtain a quality feature vector, and judges the quality feature vector according to the dynamic threshold corresponding to the fingerprint quality level. If all indicators meet the threshold, it performs afterimage detection and spur detection, and responds to the check-in request to enter the matching stage. The quality feature vector is obtained, and the quality score of the current image is obtained through the quality score regression model. The corresponding matching strategy is selected according to the quality score to perform fingerprint retrieval. If the match is successful, the attendance information is recorded. If the match fails, the user is prompted to change the fingerprint or use an alternative verification method.

[0007] Preferably, the step of performing a multi-dimensional quality assessment on the current fingerprint image to obtain a quality feature vector specifically includes: Perform effective region segmentation on the current fingerprint image to extract the region where the fingerprint is located; The image sharpness index, ridge coverage ratio, grayscale distribution skewness, and local texture complexity are calculated in parallel within the fingerprint area, and the calculation results are combined into the quality feature vector.

[0008] Preferably, the step of determining the quality feature vector based on the dynamic threshold corresponding to the fingerprint quality level specifically includes: Obtain the lower limit of clarity, the lower limit of coverage ratio, and the tolerance range of grayscale distribution skewness corresponding to the fingerprint quality level; The image sharpness index in the quality feature vector is compared with the lower limit of sharpness, the ridge coverage ratio is compared with the lower limit of coverage ratio, the gray-level distribution skewness is compared with the gray-level distribution skewness tolerance range, and the local texture complexity is compared with the preset lower limit of texture complexity. If all the indicators meet the corresponding thresholds, the image is determined to be acceptable.

[0009] Preferably, the execution of afterimage detection and prosthesis detection specifically includes: Acquire a background cache image, calculate the image similarity between the current fingerprint image and the background cache image, and if the image similarity is greater than a preset ghost image threshold, it is determined to be a ghost image and a prompt is made to clean the acquisition window; If the image is not determined to be a ghost image, the local texture features and grayscale texture statistical features of the current fingerprint image are extracted and input into a pre-trained lightweight binary classification model. If the output result is a fake, the check-in is rejected and an alarm is triggered.

[0010] Preferably, the lightweight binary classification model is trained using the following method: Real fingerprint images and spurious fingerprint images are obtained as training samples, and local texture features and grayscale texture statistical features of each sample are extracted. The local texture features and the grayscale texture statistical features are concatenated into a feature vector. The lightweight neural network model is trained using the real labels as supervision signals, and the trained model is compressed to obtain the lightweight binary classification model.

[0011] Preferably, the step of selecting a corresponding matching strategy for fingerprint retrieval based on the quality score specifically includes: When the quality score is higher than the first threshold, a one-to-many comparison of the entire database is performed using a strict matching threshold. When the quality score is between the first threshold and the second threshold, it is first compared one-to-one with the user's historical successful template. If the match fails, a one-to-many comparison of the entire database is then performed. When the quality score is lower than the second threshold, a relaxed matching threshold is enabled and compared one-to-one with the most recent successful templates. If the matching fails, a prompt is made to change the finger or use an alternative verification method.

[0012] Preferably, obtaining the quality score of the current image through a quality score regression model specifically includes: Obtain historical fingerprint images and their corresponding successful matching labels; train an initial regression model based on the quality feature vector of the historical fingerprint images and the successful matching labels to obtain the quality score regression model. The quality feature vector of the current fingerprint image is input into the trained quality score regression model, and a continuous score of 0-100 is output as the quality score.

[0013] Preferably, the step of identifying the fingerprint quality level based on the fingerprint feature data specifically includes: The average ridge width and contrast distribution are obtained from the fingerprint feature data; If the average ridge width is less than a preset width threshold and the contrast distribution is lower than a preset contrast threshold, then the fingerprint quality level is identified as a faint fingerprint level, and a corresponding clarity level is set for the faint fingerprint level, wherein the lower limit of the clarity level is lower than the normal level.

[0014] The second objective of this invention is achieved through the following technical solution: A fingerprint fast retrieval system based on image pre-screening, the fingerprint fast retrieval system based on image pre-screening includes: The original feature acquisition module is used to acquire a user registration request, acquire an original fingerprint image, and extract fingerprint feature data from the original fingerprint image. The user registration module is used to identify the fingerprint quality level based on the fingerprint feature data, set the corresponding clarity level based on the fingerprint quality level, and respond to the user registration request. The user threshold recognition module is used to obtain user check-in requests, obtain the current fingerprint image, perform multi-dimensional quality assessment on the current fingerprint image to obtain a quality feature vector, and determine the quality feature vector according to the dynamic threshold corresponding to the fingerprint quality level. If all indicators meet the threshold, the module performs afterimage detection and spur detection, and responds to the check-in request to enter the matching stage. The user verification module is used to obtain the quality feature vector, obtain the quality score of the current image through the quality score regression model, select the corresponding matching strategy to perform fingerprint retrieval based on the quality score, record attendance information if the match is successful, and prompt the user to change the fingerprint or use an alternative verification method if the match fails.

[0015] In summary, this application includes at least one of the following beneficial technical effects: 1. By identifying fingerprint quality levels and setting corresponding clarity levels based on fingerprint feature data during the registration phase, and judging the quality feature vector based on the dynamic threshold corresponding to the fingerprint quality level during the check-in phase, a differentiated admission standard for people with faint fingerprints is achieved, which improves the check-in pass rate for such users. 2. A quality feature vector is obtained through multi-dimensional quality assessment, and afterimage detection and fake image detection are performed before entering the matching stage. Low-quality and fake images are filtered from multiple dimensions, which improves the accuracy of pre-screening and the anti-fraud capability of the system. 3. By mapping the quality feature vector to continuous scores through a quality score regression model, and selecting the corresponding matching strategy for fingerprint retrieval based on the quality score, the matching process achieves dynamic adaptation to image quality, thereby improving the overall retrieval efficiency while ensuring security. Attached Figure Description

[0016] Figure 1 This is a flowchart of a fingerprint fast retrieval method based on image pre-screening in one embodiment of this application; Figure 2 This is a block diagram of a fingerprint fast retrieval system based on image pre-screening in one embodiment of this application. Detailed Implementation

[0017] The present application will be further described in detail below with reference to the accompanying drawings.

[0018] In one embodiment, such as Figure 1As shown, this application discloses a fast fingerprint retrieval method based on image pre-screening, which specifically includes the following steps: S10: Obtain user registration request, obtain original fingerprint image, and extract fingerprint feature data from original fingerprint image.

[0019] In this embodiment, the user registration request refers to an instruction triggered by the administrator on the device management interface or automatically generated by the system after the user presses the fingerprint sensor when a new employee uses the fingerprint attendance machine for the first time. The original fingerprint image refers to the grayscale image captured by the sensor when the user presses their finger on the optical or semiconductor sensor, typically with a resolution of 256×288 or 320×240 pixels. Fingerprint feature data refers to the digital information extracted from the original image used to uniquely identify the user, including but not limited to the position and orientation of details such as ridge endpoints and bifurcation points.

[0020] Specifically, when a new employee joins the company, the administrator selects "Add User" in the attendance machine's backend and prompts the employee to place a designated finger (e.g., the right index finger) smoothly on the fingerprint collection window. After detecting finger pressure, the sensor continuously acquires three frames of raw fingerprint images and selects the highest quality frame as the registration template. The system performs preprocessing on this frame: first, it extracts the central region of the fingerprint using an effective region segmentation algorithm, removing background and edge noise; then, it enhances the contrast of the fingerprint ridges using orientation field estimation and Gabor filtering; finally, it extracts the minutiae of the ridges, including endpoints and bifurcation points, and records the coordinates, angle, and type of each minutiae, forming the user's fingerprint feature data. This feature data is encrypted and stored locally on the device or in a cloud database, and associated with the user's employee ID, name, and other information. To ensure registration quality, the system prompts the user to press the fingerprint twice more at different angles, and the intersection or average of the three feature data points is used as the final template to improve the robustness of subsequent matching.

[0021] S20: Identify the fingerprint quality level based on the fingerprint feature data, set the corresponding clarity level according to the fingerprint quality level, and respond to the user registration request.

[0022] In this embodiment, fingerprint quality level refers to a category based on the physical characteristics of a user's fingerprint (such as ridge depth, ridge width, and contrast), used to distinguish between the general population and those with faint fingerprints (such as blue-collar workers and the elderly). Clarity level refers to the stringency of the fingerprint image clarity requirements during the pre-screening stage; a higher level indicates more lenient requirements.

[0023] Specifically, after acquiring a user's fingerprint feature data, the system extracts two key statistics: average ridge width and ridge-valley contrast distribution. Average ridge width is obtained by measuring the average pixel width of all ridges in the feature data; for individuals with faint fingerprints, the ridges are typically narrower and blurrier. Ridge-valley contrast distribution refers to the statistical distribution of the grayscale difference between ridges and valleys in the fingerprint image; lower contrast indicates less distinct fingerprint patterns. The system compares these two indicators with preset thresholds: if the average ridge width is less than the preset width threshold (e.g., 5 pixels) and the mean of the contrast distribution is less than the preset contrast threshold (e.g., 30 grayscale levels), the user's fingerprint quality level is identified as faint; otherwise, it is identified as normal. For users with faint fingerprints, the system automatically sets their clarity level to a lower level, i.e., lowering the lower limit of clarity for subsequent check-ins (e.g., from 50 to 30), while expanding the tolerance range for dry / wet fingerprint discrimination. This setting is written to the user's profile and serves as the basis for subsequent dynamic threshold adjustments. After completing the level identification and setting, the system responds to the user registration request, returns a registration success message, and saves the user's fingerprint template and level information to the device's memory.

[0024] S30: Obtain the user's check-in request, obtain the current fingerprint image, perform multi-dimensional quality assessment on the current fingerprint image to obtain the quality feature vector, determine the quality feature vector according to the dynamic threshold corresponding to the fingerprint quality level, and if all indicators meet the threshold, perform afterimage detection and spur detection, and respond to the check-in request to enter the matching stage.

[0025] In this embodiment, a user clock-in request refers to the recognition command triggered when an employee presses their finger on the fingerprint sensor during their arrival or departure. The current fingerprint image refers to a single-frame grayscale image captured in real-time by the sensor at the moment of the press. The quality feature vector is an array of multiple dimensional indicators used to comprehensively describe the usability of the current fingerprint image. The dynamic threshold is a pre-set judgment standard based on the user's fingerprint quality level, which can be updated according to the user profile. Image retention detection identifies whether the fingerprint imprint of the previous user remains on the acquisition window, while spurious fingerprint detection identifies whether the current press is a forged fingerprint such as a silicone film.

[0026] Specifically, when an employee places their finger on the attendance machine upon arrival, the sensor immediately captures a grayscale image. The system first segments the image into an effective region, extracting the area containing the fingerprint. Then, within this region, it calculates four quality metrics in parallel: image sharpness (obtained by calculating the variance of the grayscale gradient within the region, reflecting the image's sharpness), ridge coverage ratio (the proportion of ridge pixels to total pixels after binarization, reflecting the effective fingerprint area), grayscale distribution skewness (a skewness coefficient calculated using a histogram, used to determine the dryness or wetness of the finger), and local texture complexity (reflecting the richness of the texture patterns by analyzing the diversity of texture patterns in the local neighborhood). These four metrics are combined into a four-dimensional vector, which is the quality feature vector.

[0027] Next, the system reads the fingerprint quality level recorded during user registration and obtains the corresponding dynamic threshold parameters: lower limit for sharpness, lower limit for coverage ratio, tolerance range for grayscale distribution skewness, and lower limit for texture complexity. Each item in the quality feature vector is compared with its corresponding threshold: image sharpness must be greater than or equal to the lower limit for sharpness, ridge coverage ratio must be greater than or equal to the lower limit for coverage ratio, grayscale distribution skewness must fall within the tolerance range, and local texture complexity must be greater than or equal to the lower limit for texture complexity. Only when all four indicators meet the conditions is the current image deemed acceptable and proceeds to the next detection stage; otherwise, the system directly prompts the user to press again and records the reason for failure.

[0028] For acceptable images, the system further performs afterimage detection: it acquires the background cache image saved after the last check-in (i.e., the empty window image captured when no one pressed the button), and calculates the normalized cross-correlation coefficient between the current image and the background image. If the coefficient is greater than a preset afterimage threshold (e.g., 0.7), it determines that there is residual fingerprint on the capture window and prompts the user to wipe it and try again; if no afterimage is detected, it continues to perform spurious detection: it extracts the local texture features and grayscale texture statistical features of the current image, inputs them into a pre-trained lightweight binary classification model, and if the model output is a spurious image, it rejects the check-in and triggers an alarm; if it is real, it officially responds to the check-in request and sends the current image to the matching stage.

[0029] S40: Obtain the quality feature vector, obtain the quality score of the current image through the quality score regression model, select the corresponding matching strategy to perform fingerprint retrieval based on the quality score, record the attendance information if the match is successful, and prompt the user to change the fingerprint or use the alternative verification method if the match fails.

[0030] In this embodiment, the quality score regression model refers to a pre-trained mathematical model that maps quality feature vectors to a continuous score of 0-100, where the score represents the probability of a successful match for the current image. The matching strategy employs different comparison methods and thresholds based on the quality score to balance speed and security. Alternative verification methods include entering an employee ID password or using another registered fingerprint.

[0031] Specifically, after the images pass the pre-screening, the system inputs the quality feature vector generated in step S30 into the quality score regression model. This model is trained based on historical data and outputs a quality score. Based on the quality score, the matching module selects one of three strategies: When the quality score is higher than the first threshold (e.g., 70 points), it indicates excellent image quality. The system uses a strict matching threshold (e.g., an error acceptance rate of 0.001%) to perform a one-to-many comparison across the entire database, comparing it with all registered templates one by one to ensure high security. When the quality score is between the first and second thresholds (e.g., 40 to 70 points), the system prioritizes a one-to-one comparison with the user's historical successful templates (comparing only with the template from the user's last successful check-in). If a match is successful, the user is quickly allowed to proceed; if not, a one-to-many comparison across the entire database is performed, balancing efficiency and reliability. When the quality score is lower than the second threshold (e.g., 40 points), it indicates poor image quality. The system uses a more lenient matching threshold (e.g., an error acceptance rate of 0.01%) and only performs a one-to-one comparison with the most recent (e.g., the last three) successful templates, narrowing the search range to improve speed. If a match fails, the user is prompted to try a different finger or enter their employee ID and password as backup verification methods. Regardless of the strategy used, once a match is successful, the system records the current time, user ID, matching time, and other information as attendance records, and updates the user's historical success template; if a match fails, the system notifies the user and records an exception log for administrators to analyze later.

[0032] In this embodiment, by acquiring user registration requests and extracting fingerprint feature data from original fingerprint images, the system achieves digital acquisition of basic fingerprint information. By identifying fingerprint quality levels based on fingerprint feature data and setting corresponding clarity levels, the system can establish personalized processing baselines for users with different fingerprint conditions during the registration stage, improving the targeting of subsequent processing. By acquiring user check-in requests and performing multi-dimensional quality assessments on the current fingerprint image to obtain quality feature vectors, the system comprehensively characterizes the usability of the current image, reducing the possibility of misjudgment based on a single indicator. By judging the quality feature vectors according to the dynamic thresholds corresponding to the fingerprint quality levels, and performing ghosting detection and spoofing detection when conditions are met, low-quality and forged images are filtered out before entering the matching stage, improving the security and reliability of the system. By acquiring the quality feature vectors and obtaining the quality score of the current image through a quality score regression model, and selecting the corresponding matching strategy for fingerprint retrieval based on the quality score, the matching process can dynamically adjust the retrieval range and strictness according to image quality, improving the overall retrieval efficiency. At the same time, when matching fails, the system prompts users to change their fingerprint or use an alternative verification method, providing a reasonable fault-tolerant path and improving the user experience.

[0033] In one embodiment, step S30, which involves performing a multi-dimensional quality assessment on the current fingerprint image to obtain a quality feature vector, specifically includes: S31: Perform effective region segmentation on the current fingerprint image and extract the region where the fingerprint is located.

[0034] In this embodiment, effective region segmentation refers to the process of separating the central region containing the fingerprint ridges from the entire fingerprint grayscale image, excluding background, black borders, and noise pixels. The fingerprint region is the only valid range for calculating all subsequent quality indicators, and its accuracy directly affects the reliability of the quality assessment.

[0035] Specifically, the system first performs effective region segmentation on the raw grayscale image acquired by the sensor. The segmentation process employs an adaptive thresholding algorithm, dynamically determining the segmentation threshold based on the local grayscale distribution of the image, converting the image into a binary image where ridge regions are displayed as black, and valley regions and background regions as white. Subsequently, the system performs morphological opening operations on the binary image, using circular structuring elements to remove isolated noise and small holes, and then uses connected component analysis to find the largest connected region, which is the fingerprint region. The system uses the minimum bounding rectangle of this region as the region of interest, and all subsequent calculations are confined to this region to avoid interference from background noise in quality assessment. During segmentation, the system also considers possible finger tilt or offset, using rotation correction to make the fingerprint region as horizontally centered as possible. For individuals with faint fingerprints, due to their low ridge contrast, the adaptive thresholding algorithm may struggle to segment accurately; therefore, the system additionally employs a local variance enhancement method to first improve ridge contrast before binarization, ensuring the completeness of effective region extraction. After segmentation, the system outputs a mask image marking the pixel positions of the fingerprint region for subsequent parallel computation.

[0036] S32: Calculate the image sharpness index, ridge coverage ratio, grayscale distribution skewness, and local texture complexity in parallel within the fingerprint area, and combine the calculation results into a quality feature vector.

[0037] In this embodiment, the image sharpness index is used to quantify the sharpness of the fingerprint ridges, the ridge coverage ratio is used to assess the size of the effective fingerprint area, the grayscale distribution skewness is used to determine the dryness or wetness of the finger, and the local texture complexity is used to characterize the richness of the ridge details. The quality feature vector is a four-dimensional array composed of these four indices in a fixed order, serving as the unified input for subsequent dynamic threshold determination and quality score regression.

[0038] Specifically, the system calculates four quality metrics in parallel within the fingerprint area. The image sharpness metric is calculated as follows: the second derivative of the grayscale value of each pixel within the area is calculated using the Laplacian operator to obtain a Laplacian response map. The variance of all pixel values ​​in this response map is then calculated; a larger variance indicates sharper image edges and clearer textures. The ridge coverage ratio is calculated by calculating the proportion of black pixels (ridges) to the total pixels in the binary image. A low ratio indicates that the finger is pressed too far to the correct area or the pressing area is insufficient; a normal pressing area should be above 0.5. The grayscale distribution skewness is calculated by calculating the skewness coefficient of the grayscale histogram within the area. A positive skewness value indicates that the overall image is too bright, possibly indicating a dry finger; a negative skewness value indicates that the overall image is too dark, possibly indicating a wet finger; a smaller absolute value indicates a moderate dry / wet condition. The method for calculating local texture complexity is as follows: the region is divided into several fixed-size sub-blocks, and a histogram of local binary patterns is calculated within each sub-block. Then, the information entropy is calculated for the histograms of all sub-blocks. The higher the entropy value, the richer the texture details and the more complex the texture. After the four indicators are calculated, the system combines them into a four-dimensional vector in the order of sharpness, coverage ratio, skewness, and texture complexity, which is the quality feature vector. This vector serves as the input for subsequent steps, used for dynamic threshold determination and quality score regression.

[0039] In this embodiment, by performing effective region segmentation on the current fingerprint image and extracting the fingerprint area, the interference of background and invalid regions on subsequent analysis is eliminated, improving the accuracy of quality assessment. By parallelly calculating image sharpness index, ridge coverage ratio, gray-level distribution skewness, and local texture complexity within the fingerprint area, the usability of the fingerprint image is evaluated from multiple complementary dimensions. Among them, the image sharpness index reflects the sharpness of the image, the ridge coverage ratio reflects the size of the effective fingerprint area, the gray-level distribution skewness helps to determine the dryness or wetness of the finger, and the local texture complexity characterizes the richness of the texture. The combination of these four indicators forms a comprehensive quantitative description of the fingerprint image quality, reducing misjudgments caused by the deviation of a single indicator. By combining the calculation results into a quality feature vector, a structured input is provided for subsequent dynamic threshold determination and quality score regression, improving the stability and reproducibility of the entire pre-screening process.

[0040] In one embodiment, step S30, which involves determining the quality feature vector based on a dynamic threshold corresponding to the fingerprint quality level, specifically includes: S33: Obtain the lower limit of clarity, the lower limit of coverage ratio, and the tolerance range of grayscale distribution skewness corresponding to the fingerprint quality level.

[0041] In this embodiment, the lower limit of sharpness is the minimum allowable value of the image sharpness index, the lower limit of coverage ratio is the minimum allowable value of the ridge coverage ratio, and the grayscale distribution skewness tolerance range is the allowable range of skewness. These three parameters together constitute the pre-screening threshold for a specific fingerprint quality level. The fingerprint quality level is a classification determined based on the user's fingerprint characteristics during the registration stage, and different levels correspond to different parameter values.

[0042] Specifically, the system reads the fingerprint quality level from the user's profile. If the level is "faint fingerprint," the system loads the corresponding threshold parameters from the device's memory: a minimum clarity of 30, a minimum coverage ratio of 0.4, and a grayscale distribution skew tolerance range from -1.5 to +1.5. If the level is "normal," the system loads the default, higher-requirement parameters: a minimum clarity of 50, a minimum coverage ratio of 0.6, and a grayscale distribution skew tolerance range from -1.0 to +1.0. These parameters are default values ​​set at the factory based on statistical analysis of a large number of samples and can be fine-tuned by the administrator through the backend based on actual usage. After obtaining the parameters, the system temporarily stores them in memory, waiting to compare them with the quality feature vector. It is important to note that the grayscale distribution skew tolerance range is a range rather than a single threshold. This means that as long as the skew falls within this range, whether positive or negative, it is considered acceptable, thus accommodating a certain degree of dry and wet fluctuations.

[0043] S34: Compare the image sharpness index in the quality feature vector with the sharpness lower limit, the ridge coverage ratio with the coverage ratio lower limit, the gray-level distribution skewness with the gray-level distribution skewness tolerance range, and the local texture complexity with the preset texture complexity lower limit. If all indicators meet the corresponding thresholds, the image is determined to be acceptable.

[0044] In this embodiment, the comparison process involves comparing each component of the quality feature vector with its corresponding threshold one by one. Only when all components meet the conditions is the current image deemed acceptable, allowing it to proceed to the subsequent ghosting detection and spurious detection stages. The preset texture complexity lower limit is a fixed value independent of the fingerprint quality level, used to ensure that the image has sufficient texture detail.

[0045] Specifically, the system compares the four components of the quality feature vector with their corresponding thresholds sequentially. First, it compares the image sharpness index with the lower limit of sharpness. If the sharpness index is lower than the lower limit, the image is considered blurry, and a prompt to press again is returned. Second, it compares the ridge coverage ratio with the lower limit of coverage ratio. If the coverage ratio is lower than the lower limit, the effective area is considered insufficient, and a prompt to press again is returned. Third, it compares the grayscale distribution skewness with the tolerance range. If the skewness is lower than the lower limit or higher than the upper limit of the range, the dry / wet state is considered abnormal, and a prompt to press again is returned. Finally, it compares the local texture complexity with the preset lower limit of texture complexity. If the complexity is lower than the lower limit, the texture is considered too simple, possibly indicating a fake or improper pressing, and a prompt to press again is returned. Only when all four conditions are met does the system determine the current image as acceptable and output a pass flag. If any condition is not met, the system, while returning a prompt to press again, will record the specific failure dimension of this judgment in the log, such as insufficient sharpness or low coverage ratio, so as to optimize the threshold or remind the user to adjust the pressing method later. This multi-dimensional joint judgment method ensures that users with shallow fingerprints will not be directly rejected due to their low clarity caused by shallow fingerprint patterns. As long as other dimensions meet the standards, they still have a chance to pass, thereby improving the pass rate of such users.

[0046] In this embodiment, by obtaining the lower limit of clarity, the lower limit of coverage ratio, and the tolerance range of grayscale distribution skewness corresponding to the fingerprint quality level, differentiated judgment criteria for users with different fingerprint conditions are realized, allowing people with faint fingerprints to obtain more lenient admission conditions. By comparing the image clarity index in the quality feature vector with the lower limit of clarity, the ridge coverage ratio with the lower limit of coverage ratio, the grayscale distribution skewness with the tolerance range of grayscale distribution skewness, and the local texture complexity with the preset lower limit of texture complexity, the current image is verified from four key dimensions to ensure that the judgment process is comprehensive and rigorous. By only judging an image as acceptable when all indicators meet the corresponding thresholds, the situation of mistakenly accepting low-quality images due to abnormal individual indicators is avoided. While relaxing the admission conditions, the basic image quality standard is maintained, improving the pass rate of people with faint fingerprints and reducing recognition failures caused by insufficient image quality in the subsequent matching stage.

[0047] In one embodiment, step S30, namely performing afterimage detection and prosthesis detection, specifically includes: S35: Obtain the background cache image, calculate the image similarity between the current fingerprint image and the background cache image. If the image similarity is greater than the preset afterimage threshold, it is determined to be an afterimage image, and a prompt to clean the acquisition window is given.

[0048] In this embodiment, the background cache image refers to the blank image of the collection window periodically collected by the attendance machine when no one is pressing the button, which is used as a reference benchmark for afterimage detection. Image similarity is measured by calculating the normalized cross-correlation coefficient between two images, with a value ranging from -1 to +1. The closer the value is to +1, the more similar the two images are. The preset afterimage threshold is the critical value for determining whether there is residual fingerprint, usually set to 0.7. Afterimage image refers to the situation where the fingerprint imprint of the previous user remains on the collection window, resulting in superimposed extra lines in the current image.

[0049] Specifically, the attendance machine automatically captures a background image and stores it in a cache after each successful attendance check-in. The background cache is also automatically updated every five minutes to adapt to slow changes in ambient light. When a user presses the fingerprint sensor, the system reads the latest background image from the cache and calculates the normalized cross-correlation coefficient between the current fingerprint image and the background image. During the calculation, the system scales both images to the same size, standardizes the grayscale value of each corresponding pixel, and then calculates the correlation coefficient. If the coefficient is greater than a preset afterimage threshold of 0.7, it is determined that there is residual fingerprint on the collection window. The system displays a prompt on the screen and issues a voice reminder, asking the user to wipe the collection window with a clean cloth and press again. If the coefficient does not exceed 0.7, it is determined that there is no afterimage, and proceeds to the next sub-step. The purpose of afterimage detection is to prevent the fingerprint of the previous user from being mistakenly identified as the current user's fingerprint, thereby avoiding incorrect identification or matching failure.

[0050] S36: If the image is not identified as a ghost image, extract the local texture features and grayscale texture statistical features of the current fingerprint image and input them into the pre-trained lightweight binary classification model. If the output result is a fake, refuse to clock in and trigger an alarm.

[0051] In this embodiment, local texture features refer to descriptors reflecting local texture patterns extracted from the fingerprint image, and grayscale texture statistical features refer to parameters reflecting texture statistical characteristics extracted from the grayscale co-occurrence matrix. The lightweight binary classification model is a small, compressed neural network capable of quickly outputting a determination of whether the current image belongs to a real finger or a fake one. Fake fingerprints include silicone films, gelatin films, and other forged fingerprints.

[0052] Specifically, the system extracts local texture features from the effective area of ​​the current fingerprint image. The extraction method involves dividing the region into several 16x16 pixel sub-blocks, calculating the histogram of local binary patterns within each sub-block, and then concatenating the histograms of all sub-blocks into a long vector. Simultaneously, the system extracts four statistical measures from the gray-level co-occurrence matrix as gray-level texture statistical features: contrast, correlation, energy, and homogeneity. The local texture feature vector and the gray-level texture statistical feature vector are concatenated into a complete feature vector, which is then input into a pre-trained lightweight binary classification model. The model's forward propagation outputs a probability value between 0 and 1. If the probability value is greater than 0.5, the model determines the current image is a fake, the system immediately rejects the attendance attempt, triggers an audible and visual alarm, and records the event details in the log for security personnel to verify. If the probability value is less than or equal to 0.5, it is determined to be a genuine fingerprint, and the current image is allowed to proceed to the matching stage. Fake fingerprint detection, as a second line of defense, effectively intercepts malicious attendance attempts using forged fingerprints, improving the security of the attendance system.

[0053] In this embodiment, by acquiring a background cache image and calculating the image similarity between the current fingerprint image and the background cache image, residual fingerprint traces left by previous users on the collection window can be effectively identified. When the similarity is greater than a preset afterimage threshold, it is determined to be an afterimage image and a prompt to clean the collection window is given. This reduces the occurrence of incorrect recognition and repeated check-in failures caused by physical stains, improving the hygiene level and recognition accuracy of the device. After the current fingerprint image is not determined to be an afterimage image, local texture features and grayscale texture statistical features are extracted and input into a pre-trained lightweight binary classification model. Texture statistical differences are used to distinguish between real fingerprints and fakes such as silicone films. When the output result is a fake, check-in is rejected and an alarm is triggered. Forgery attacks are intercepted in the pre-screening stage, enhancing the system's anti-fraud capabilities. Afterimage detection and fake detection serve as two front-line defenses, improving overall security without affecting the check-in speed of normal users.

[0054] In one embodiment, in step S36, the lightweight binary classification model is trained using the following method: S361: Obtain real fingerprint images and spurious fingerprint images as training samples, and extract local texture features and grayscale texture statistical features for each sample.

[0055] In this embodiment, real fingerprint images refer to actual finger pressing images collected from people of different ages, genders, and occupations, covering various states such as dry / wet and deep / shallow. Fake fingerprint images refer to forged fingerprint pressing images made using materials such as silicone, gelatin, and latex. The quantity and quality of training samples directly affect the model's generalization ability. Local texture features and grayscale texture statistical features together constitute the discriminative description that distinguishes real from fake fingerprints.

[0056] Specifically, a training dataset was first constructed. Over 10,000 real fingerprint images were collected from the daily operation of the attendance machine, covering samples from different fingers, different pressure levels, and different environmental conditions. Simultaneously, prosthetic fingerprints were fabricated using commercially available materials, and over 10,000 prosthetic images were collected on the same sensor to ensure a balanced number of samples from both classes. For each image, effective region segmentation was first performed to extract the fingerprint region. Then, local texture features were extracted within this region. Specifically, the region was divided into 16x16 pixel sub-blocks, and a histogram of the local binary pattern was calculated within each sub-block. The histograms of all sub-blocks were concatenated into a 256-dimensional feature vector. Simultaneously, four statistics were extracted from the gray-level co-occurrence matrix: contrast reflects the depth of the texture, correlation reflects the directional consistency of the texture, energy reflects the uniformity of the texture, and homogeneity reflects the smoothness of the texture. These four statistics formed a 4-dimensional gray-level texture statistical feature vector. The local texture feature vector and the gray-level texture statistical feature vector were concatenated into a 260-dimensional feature vector as the model input. Each sample is labeled with a real label: a real finger is labeled as 1, and a prosthetic finger is labeled as 0.

[0057] S362: Local texture features and grayscale texture statistical features are concatenated into a feature vector. The lightweight neural network model is trained using the real labels as supervision signals, and the trained model is compressed to obtain a lightweight binary classification model.

[0058] In this embodiment, the lightweight neural network model is a network structure with few parameters and low computational cost, suitable for deployment on resource-constrained embedded devices. Compression processing includes pruning and quantization operations to reduce model size and inference time, enabling it to run in real-time on an MCU.

[0059] Specifically, a lightweight neural network model was designed, comprising an input layer with a dimension of 260, a fully connected hidden layer with 64 neurons using the ReLU activation function, and an output layer with 2 neurons using the Softmax activation function for binary classification. Training was performed using the cross-entropy loss function and the Adam optimizer, with a batch size of 32 and 50 training iterations. During training, the model continuously adjusted the weights through backpropagation, ensuring that the output probability for real fingerprints was close to 1 and the output probability for spoofed fingerprints was close to 0. After training, the model was compressed: first, weight pruning was performed, resetting connection weights with an absolute value less than 0.01 to zero and removing approximately 30% of redundant connections; then, int8 quantization was performed, converting all weights and activation values ​​from 32-bit floating-point numbers to 8-bit integers. The compressed model size was less than 200KB, and the single inference time on a 200MHz MCU did not exceed 10 milliseconds. After being deployed to the attendance machine, the model can distinguish between real fingerprints and common fake fingerprints in real time, and has a high recognition rate for real fingerprints of people with light fingerprints, and will not be misjudged as fakes due to light fingerprints.

[0060] In this embodiment, by acquiring real fingerprint images and spurious fingerprint images as training samples, the coverage and representativeness of the training data are ensured, enabling the model to learn the differences between real and spurious fingerprints in local texture features and grayscale texture statistical features. By extracting the local texture features and grayscale texture statistical features of each sample and concatenating them into a feature vector, the high-dimensional information of the original image is compressed into low-dimensional discriminative features, reducing the computational complexity of the model. By training the lightweight neural network model with real labels as supervision signals, the model can autonomously fit the nonlinear mapping relationship between features and categories, improving the accuracy of spurious fingerprint recognition. By compressing the trained model, a small-sized lightweight binary classification model is obtained, which can be deployed on resource-constrained embedded devices, meeting real-time requirements while maintaining high recognition performance, thus improving the engineering practicality of the method.

[0061] In one embodiment, step S40, which involves selecting a corresponding matching strategy based on the quality score for fingerprint retrieval, specifically includes: S41: When the quality score is higher than the first threshold, a one-to-many comparison of the entire database is performed using a strict matching threshold.

[0062] In this embodiment, the first threshold is a preset quality score threshold, typically set to 70. A strict matching threshold refers to a matching standard that requires a large number of matching feature points with small positional errors, used to ensure high security. A one-to-many comparison across the entire database refers to comparing the current fingerprint with each of all registered user templates stored in the device.

[0063] Specifically, when the quality score is greater than 70, it indicates that the current fingerprint image quality is excellent, and the system adopts a strict matching threshold. The strict matching threshold requires that the number of matched minutiae be greater than or equal to 15, and the positional error of the corresponding minutiae be less than or equal to 3 pixels, and the directional error be less than or equal to 15 degrees. The system compares the feature data of the current fingerprint with the templates of all registered users in the database one by one, calculating the matching score for each pair of templates. If a user's matching score exceeds the strict threshold, the match is considered successful, and the user's attendance information is recorded; if the matching scores of all users do not exceed the threshold, the match is considered unsuccessful, and the user is prompted to retry. This strategy pursues ultimate security when the image quality is high, and is suitable for rapid release of high-confidence images during peak hours, while avoiding the risk of false identification due to excessively low thresholds.

[0064] S42: When the quality score is between the first threshold and the second threshold, a one-to-one comparison with the user's historical successful templates is performed first. If the match fails, a one-to-many comparison of the entire database is then performed.

[0065] In this embodiment, the second threshold is another quality score threshold below the first threshold, typically set to 40. The historical successful template refers to the fingerprint template saved during the user's most recent successful check-in. One-to-one comparison means comparing only the current fingerprint with the template of a specific user, rather than searching the entire database.

[0066] Specifically, a quality score between 40 and 70 indicates medium image quality. The system first retrieves the fingerprint template saved from the user's most recent successful check-in from the database and performs a one-to-one comparison using the default matching threshold (12 or more minutiae and a positional error of 4 pixels or less). If a match is successful, the user is directly approved and attendance information is recorded without a full database search, significantly reducing matching time. If a match fails, it is downgraded to a full database one-to-many comparison, using the default matching threshold to compare all registered templates one by one, ensuring that there are no false rejections due to expired historical templates or recent changes in the user's fingerprint. This strategy balances efficiency and accuracy, as the user's own template is sufficient for a successful match in most cases, and a full database search is only performed when a template becomes invalid.

[0067] S43: When the quality score is lower than the second threshold, enable the lenient matching threshold and compare it one-to-one with the most recent successful templates. If the matching fails, prompt to change the finger or use the alternative verification method.

[0068] In this embodiment, the relaxed matching threshold refers to a matching standard that appropriately reduces the required number of matches and the tolerance for positional errors. The "most recent successful templates" refer to the fingerprint templates saved during the user's three most recent successful check-ins. Alternative verification methods include entering the employee ID password or using another registered fingerprint.

[0069] Specifically, a quality score below 40 indicates poor image quality. The system uses a lenient matching threshold, requiring at least 8 matching detail points, a positional error of at least 5 pixels, and an orientation error of at least 25 degrees. The system only compares the image with the user's three most recent successful templates, using the lenient threshold for each comparison. A successful match in any of the three comparisons is considered acceptable. This strategy of narrowing the search scope improves the matching speed for low-quality images and avoids long wait times caused by full-database searches. If all three comparisons fail, the system prompts the user to try pressing with a different finger or enter their employee ID and password as backup verification methods. Simultaneously, the system records this low-quality score event, including the user's employee ID, quality score, and matching result, for subsequent analysis to determine if adjusting the user's dynamic threshold or notifying the administrator to check the sensor status is necessary.

[0070] In this embodiment, when the quality score is higher than the first threshold, a strict matching threshold is used to perform a one-to-many comparison across the entire database, ensuring high security for high-quality images and reducing the occurrence of erroneous acceptances. When the quality score is between the first and second thresholds, a one-to-one comparison with the user's historical successful templates is prioritized, using the user's past successful records to quickly verify identity, reducing the time overhead of the entire database search and improving the retrieval speed for medium-quality images. A one-to-many comparison across the entire database is only performed as a fallback when a match fails, balancing efficiency and reliability. When the quality score is lower than the second threshold, a relaxed matching threshold is enabled, and a one-to-one comparison is performed with the most recent successful templates, providing a limited search range for low-quality images, reducing the time consumption and risk of false rejections caused by poor image quality in the entire database search. If a match fails, the user is prompted to change their fingerprint or use an alternative verification method, providing clear guidance for the next step and improving the system's fault tolerance and user experience.

[0071] In one embodiment, step S40, which involves obtaining the quality score of the current image through a quality score regression model, specifically includes: S44: Obtain historical fingerprint images and their corresponding successful matching labels. Train the initial regression model based on the quality feature vector of the historical fingerprint images and the successful matching labels to obtain the quality score regression model.

[0072] In this embodiment, historical fingerprint images refer to a large number of attendance images accumulated from the daily operation of the attendance machine, including successful and unsuccessful matching cases. A successful match label indicates whether the image successfully found the corresponding user in subsequent matching stages; success is recorded as 1, and failure as 0. The initial regression model is a mathematical model that can learn the mapping relationship between input features and output scores; it can be a linear regression model or a small neural network.

[0073] Specifically, firstly, all attendance records from the past six months are extracted from the attendance machine's log database. Each record contains a four-dimensional quality feature vector generated from a fingerprint image during the pre-screening stage, along with a label indicating whether the image ultimately matched successfully. At least 50,000 samples are collected to ensure statistical significance, with the ratio of successfully matched to unmatched samples being as balanced as possible to avoid model bias towards the majority class. Next, a regression model is designed with a four-dimensional input layer corresponding to the four components of the quality feature vector. The model includes a fully connected hidden layer with 32 neurons, using the ReLU activation function to introduce non-linearity, and a single neuron in the output layer, using the Sigmoid activation function to map the output to between 0 and 1. During training, the successfully matched label is used as a supervision signal, mean squared error is used as the loss function, the Adam optimizer is selected, the batch size is 64, and the training is iterated for 100 epochs. The training objective is to make the model's output score as close as possible to the probability of a successful image match. After training, the model's weights and structure are saved, resulting in the quality score regression model. The model can be trained offline using GPU acceleration, and after training, it can be exported in a lightweight format and deployed to attendance machines.

[0074] S45: Input the quality feature vector of the current fingerprint image into the trained quality score regression model, and output a continuous score of 0-100 as the quality score.

[0075] In this embodiment, the quality feature vector of the current fingerprint image consists of four dimensions generated in step S31. The quality score regression model is the model trained in sub-step 7-1. The quality score is a continuous score output by the model, representing the predicted probability of a successful match in the subsequent stages of the current image.

[0076] Specifically, when a user clocks in, the system, after completing a multi-dimensional quality assessment and obtaining a quality feature vector, directly inputs this four-dimensional vector into the quality score regression model deployed on the attendance machine. The model performs one forward propagation: the input layer receives four values, which are then weighted and summed with ReLU activation in the hidden layer, and then weighted and summed with Sigmoid activation in the output layer, finally outputting a floating-point number between 0 and 1. The system multiplies this floating-point number by 100 to obtain an integer score between 0 and 100 as the quality score. For example, if the model outputs 0.85, the quality score is 85. This quality score not only reflects the image's clarity, coverage, wetness / dryness, and texture richness, but also implicitly includes a correlation with historical matching success rates, thus possessing greater predictive power than simple rule-based scoring. For individuals with faint fingerprints, although their images have lower clarity, if the historical matching success rate for such images has been acceptable, the model may give a moderately high score, such as between 50 and 65, thus avoiding premature elimination. The quality score is then passed to the matching module for selecting the matching strategy described in claim 6.

[0077] In this embodiment, by acquiring historical fingerprint images and their corresponding successful matching labels, empirical data reflecting the correlation between image quality and actual recognition results is accumulated, providing supervised learning signals for model training. By training the initial regression model based on the quality feature vectors of historical fingerprint images and successful matching labels, the model can autonomously learn the nonlinear mapping relationship between each dimension of the quality feature vector and the matching success rate, improving the accuracy and generalization ability of quality score prediction. By inputting the quality feature vector of the current fingerprint image into the trained quality score regression model, a continuous score of 0-100 is output as the quality score, transforming the multi-dimensional qualitative evaluation results into a single numerical quantitative measure, which facilitates the fine-grained classification of subsequent matching strategies. The positive correlation design between this continuous score and the probability of successful matching ensures that high-scoring images correspond to higher credibility and low-scoring images correspond to lower credibility, providing a reliable numerical basis for the decision of the three-level matching strategy, reducing mismatches or omissions caused by inaccurate quality assessments. At the same time, the training process can be completed in an offline environment without affecting the real-time performance of online check-in.

[0078] In one embodiment, step S20, namely identifying the fingerprint quality level based on fingerprint feature data, specifically includes: S21: Obtain the average ridge width and contrast distribution from the fingerprint feature data.

[0079] In this embodiment, the average ridge width refers to the average width of all ridge pixels in the fingerprint image, expressed in pixels, reflecting the thickness of the fingerprint ridges. Contrast distribution refers to the statistical distribution of the grayscale difference between ridges and valleys in the fingerprint image; its mean is usually taken as the representative value, reflecting the degree of brightness difference in the ridges. These two parameters are key indicators for determining whether a fingerprint is of a light or faint type.

[0080] Specifically, during the user registration phase, the system calculates two key parameters from the extracted fingerprint feature data. The average ridge width is calculated as follows: a distance transformation is performed on the thinned ridge skeleton to obtain the distance value from each ridge pixel to the nearest background pixel. The average of all ridge pixel distance values ​​is then multiplied by 2 to obtain an approximate ridge width. The contrast distribution is calculated as follows: within the effective area of ​​the original grayscale image, the average grayscale value of pixels belonging to the ridges and the average grayscale value of pixels belonging to the valleys are statistically analyzed, and the absolute value of the difference between the two is used as the contrast value. The system temporarily stores these two values ​​for the next step of grade recognition.

[0081] S22: If the average ridge width is less than the preset width threshold and the contrast distribution is lower than the preset contrast threshold, the fingerprint quality level is identified as the fingerprint faint level, and a corresponding clarity level is set for the fingerprint faint level, wherein the lower limit of the clarity level is lower than the normal level.

[0082] In this embodiment, the preset width threshold and preset contrast threshold are empirical values ​​derived from statistical analysis of a large number of samples, used to distinguish between ordinary fingerprints and faint fingerprints. The faint fingerprint level is a classification specifically set for blue-collar workers, the elderly, and other groups whose fingerprints are worn or have unclear patterns. The clarity level refers to the stringency of the image clarity requirements in the pre-screening stage; the higher the level, the more lenient the requirements.

[0083] Specifically, the system compares the calculated average ridge width with a preset width threshold of 5 pixels and the contrast value with a preset contrast threshold of 30 gray levels. If the average ridge width is less than 5 pixels and the contrast value is less than 30, the user is classified as having a faint fingerprint; otherwise, they are classified as having a normal fingerprint. For users with faint fingerprints, the system automatically sets their clarity level to a low level, i.e., the lower limit of clarity is adjusted to 30, the lower limit of coverage ratio is adjusted to 0.4, the tolerance range for grayscale distribution skewness is expanded to -1.5 to +1.5, and the lower limit of texture complexity is adjusted to 0.5. These parameters are written into the user's profile and stored along with the fingerprint template. During each subsequent check-in, the system reads the user's clarity level parameters and applies them to the dynamic threshold determination in the pre-screening process. In addition, the system can dynamically fine-tune these parameters based on the user's subsequent check-in success rate. For example, if five consecutive check-ins are successful, the threshold can be tightened appropriately to maintain security; if three consecutive check-ins fail due to pre-screening and require re-pressing, the threshold can be further relaxed to improve the pass rate. This level recognition and parameter configuration based on individual characteristics eliminates the need for people with faint fingerprints to endure frequent re-press prompts, significantly improving the inclusivity and user experience of the attendance system.

[0084] In this embodiment, by obtaining the average ridge width and contrast distribution from fingerprint feature data, two core characteristics of people with faint fingerprints are captured: narrow ridges result in indistinct patterns, and low contrast results in weak differences between ridge valleys. When the average ridge width is less than a preset width threshold and the contrast distribution is lower than a preset contrast threshold, the fingerprint quality level is identified as a faint fingerprint level, achieving automatic classification of special groups such as blue-collar workers and the elderly without manual intervention. By setting a corresponding clarity level for the faint fingerprint level and making the lower limit of the clarity level lower than the normal level, a more lenient image admission standard is established for this type of user during the registration stage. This significantly reduces the probability of being falsely rejected due to faint fingerprints during subsequent check-ins, improving the check-in pass rate and satisfaction of this type of user, while maintaining the original standards for ordinary users.

[0085] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0086] In one embodiment, a fingerprint fast retrieval system based on image pre-screening is provided, which corresponds one-to-one with the fingerprint fast retrieval method based on image pre-screening described in the above embodiments. For example... Figure 2 As shown, this image-based fingerprint fast retrieval system includes an original feature acquisition module, a user registration module, a user threshold recognition module, and a user verification module. Detailed descriptions of each functional module are as follows: The original feature acquisition module is used to obtain user registration requests, acquire original fingerprint images, and extract fingerprint feature data from the original fingerprint images; The user registration module is used to identify the fingerprint quality level based on fingerprint feature data, set the corresponding clarity level according to the fingerprint quality level, and respond to user registration requests. The user threshold recognition module is used to obtain user check-in requests, acquire the current fingerprint image, perform multi-dimensional quality assessment on the current fingerprint image to obtain quality feature vectors, and determine the quality feature vectors according to the dynamic thresholds corresponding to the fingerprint quality level. If all indicators meet the thresholds, the module performs afterimage detection and spur detection, and responds to the check-in request to enter the matching stage. The user verification module is used to obtain quality feature vectors, obtain the quality score of the current image through a quality score regression model, select the corresponding matching strategy for fingerprint retrieval based on the quality score, record attendance information if the match is successful, and prompt the user to change the fingerprint or use an alternative verification method if the match fails.

[0087] Specific limitations regarding the image pre-screening-based fingerprint rapid retrieval system can be found in the limitations of the image pre-screening-based fingerprint rapid retrieval method described above, and will not be repeated here. Each module in the aforementioned image pre-screening-based fingerprint rapid retrieval system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0088] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A rapid fingerprint retrieval method based on image pre-screening, characterized in that, The fingerprint fast retrieval method based on image pre-screening includes: Obtain a user registration request, obtain an original fingerprint image, and extract fingerprint feature data from the original fingerprint image; The fingerprint quality level is identified based on the fingerprint feature data, a corresponding clarity level is set based on the fingerprint quality level, and the user registration request is responded to. The system obtains a user's check-in request, acquires the current fingerprint image, performs a multi-dimensional quality assessment on the current fingerprint image to obtain a quality feature vector, and judges the quality feature vector according to the dynamic threshold corresponding to the fingerprint quality level. If all indicators meet the threshold, it performs afterimage detection and spur detection, and responds to the check-in request to enter the matching stage. The quality feature vector is obtained, and the quality score of the current image is obtained through the quality score regression model. The corresponding matching strategy is selected according to the quality score to perform fingerprint retrieval. If the match is successful, the attendance information is recorded. If the match fails, the user is prompted to change the fingerprint or use an alternative verification method.

2. The fingerprint fast retrieval method based on image pre-screening according to claim 1, characterized in that, The process of performing a multi-dimensional quality assessment on the current fingerprint image to obtain a quality feature vector specifically includes: Perform effective region segmentation on the current fingerprint image to extract the region where the fingerprint is located; The image sharpness index, ridge coverage ratio, grayscale distribution skewness, and local texture complexity are calculated in parallel within the fingerprint area, and the calculation results are combined into the quality feature vector.

3. The fingerprint fast retrieval method based on image pre-screening according to claim 2, characterized in that, The step of determining the quality feature vector based on the dynamic threshold corresponding to the fingerprint quality level specifically includes: Obtain the lower limit of clarity, the lower limit of coverage ratio, and the tolerance range of grayscale distribution skewness corresponding to the fingerprint quality level; The image sharpness index in the quality feature vector is compared with the lower limit of sharpness, the ridge coverage ratio is compared with the lower limit of coverage ratio, the gray-level distribution skewness is compared with the gray-level distribution skewness tolerance range, and the local texture complexity is compared with the preset lower limit of texture complexity. If all the indicators meet the corresponding thresholds, the image is determined to be acceptable.

4. The fingerprint fast retrieval method based on image pre-screening according to claim 3, characterized in that, The execution of afterimage detection and prosthesis detection specifically includes: Acquire a background cache image, calculate the image similarity between the current fingerprint image and the background cache image, and if the image similarity is greater than a preset ghost image threshold, it is determined to be a ghost image and a prompt is made to clean the acquisition window; If the image is not determined to be a ghost image, the local texture features and grayscale texture statistical features of the current fingerprint image are extracted and input into a pre-trained lightweight binary classification model. If the output result is a fake, the check-in is rejected and an alarm is triggered.

5. The fingerprint fast retrieval method based on image pre-screening according to claim 4, characterized in that, The lightweight binary classification model is trained using the following method: Real fingerprint images and spurious fingerprint images are obtained as training samples, and local texture features and grayscale texture statistical features of each sample are extracted. The local texture features and the grayscale texture statistical features are concatenated into a feature vector. The lightweight neural network model is trained using the real labels as supervision signals, and the trained model is compressed to obtain the lightweight binary classification model.

6. The fingerprint fast retrieval method based on image pre-screening according to claim 1, characterized in that, The step of selecting a corresponding matching strategy based on the quality score for fingerprint retrieval specifically includes: When the quality score is higher than the first threshold, a one-to-many comparison of the entire database is performed using a strict matching threshold. When the quality score is between the first threshold and the second threshold, it is first compared one-to-one with the user's historical successful template. If the match fails, a one-to-many comparison of the entire database is then performed. When the quality score is lower than the second threshold, a relaxed matching threshold is enabled and compared one-to-one with the most recent successful templates. If the matching fails, a prompt is made to change the finger or use an alternative verification method.

7. The fingerprint fast retrieval method based on image pre-screening according to claim 6, characterized in that, The process of obtaining the quality score of the current image through a quality score regression model specifically includes: Obtain historical fingerprint images and their corresponding successful matching labels; train an initial regression model based on the quality feature vector of the historical fingerprint images and the successful matching labels to obtain the quality score regression model. The quality feature vector of the current fingerprint image is input into the trained quality score regression model, and a continuous score of 0-100 is output as the quality score.

8. The fingerprint fast retrieval method based on image pre-screening according to claim 1, characterized in that, The step of identifying the fingerprint quality level based on the fingerprint feature data specifically includes: The average ridge width and contrast distribution are obtained from the fingerprint feature data; If the average ridge width is less than a preset width threshold and the contrast distribution is lower than a preset contrast threshold, then the fingerprint quality level is identified as a faint fingerprint level, and a corresponding clarity level is set for the faint fingerprint level, wherein the lower limit of the clarity level is lower than the normal level.

9. A fingerprint rapid retrieval system based on image pre-screening, characterized in that, The fingerprint fast retrieval system based on image pre-screening includes: The original feature acquisition module is used to acquire a user registration request, acquire an original fingerprint image, and extract fingerprint feature data from the original fingerprint image. The user registration module is used to identify the fingerprint quality level based on the fingerprint feature data, set the corresponding clarity level based on the fingerprint quality level, and respond to the user registration request. The user threshold recognition module is used to obtain user check-in requests, obtain the current fingerprint image, perform multi-dimensional quality assessment on the current fingerprint image to obtain a quality feature vector, and determine the quality feature vector according to the dynamic threshold corresponding to the fingerprint quality level. If all indicators meet the threshold, the module performs afterimage detection and spur detection, and responds to the check-in request to enter the matching stage. The user verification module is used to obtain the quality feature vector, obtain the quality score of the current image through the quality score regression model, select the corresponding matching strategy to perform fingerprint retrieval based on the quality score, record attendance information if the match is successful, and prompt the user to change the fingerprint or use an alternative verification method if the match fails.

Images