A partial fingerprint recognition method based on multi-level feature fusion

Abstract

This invention discloses a partial fingerprint recognition method based on multi-level feature fusion, relating to the fields of image processing and pattern recognition. The method includes: preprocessing an input fingerprint image to obtain a preprocessed fingerprint image; extracting primary curvature features, secondary minutiae features, and tertiary ridge shape features based on the preprocessed fingerprint image; retrieving and classifying a fingerprint template library using the primary curvature features to obtain a candidate fingerprint set; performing minutiae matching and ridge shape matching based on the candidate fingerprint set using the secondary minutiae features and tertiary ridge shape features to obtain minutiae matching scores and ridge shape matching scores; and calculating the final matching score using adaptive weighted fusion to obtain the fingerprint matching determination result. By fusing multi-level features such as curvature, minutiae, and ridge shape, and employing an adaptive weighting strategy, this method improves the efficiency, accuracy, and adaptability to images of varying quality in partial fingerprint recognition.

Description

Technical Field

[0001] This invention relates to the field of image processing and pattern recognition, and more specifically to a partial fingerprint recognition method based on multi-level feature fusion. Background Technology

[0002] Fingerprint recognition technology, especially in fields such as criminal investigation, access control and security, and mobile device authentication, often needs to process incomplete or low-quality fingerprint images, i.e. partial fingerprints. Due to factors such as incomplete pressing, sensor area limitations, or conditions left at the scene, these images often only contain a local area of ​​the fingerprint, and their feature information is limited, which poses a challenge to accurate individual identification.

[0003] In existing technologies, a common approach is to rely on a single type of feature for identification. For example, some methods match only based on minutiae, such as ridge endpoints and bifurcation points. However, in some fingerprint scenarios, due to the small effective area, the number of minutiae that can be extracted may be severely insufficient, resulting in inadequate feature representation and a sharp decline in matching reliability. Another approach is to use only macroscopic orientation field or curvature features for classification and retrieval. Although these features are computationally efficient, their discriminative power is limited, making it impossible to independently complete high-precision identity verification and difficult to achieve final accurate matching after narrowing down the candidate range.

[0004] To improve recognition performance, some improvement schemes attempt to combine multiple features. However, when fusing features at different levels, existing methods mostly use fixed fusion rules or weights. This static fusion strategy lacks flexibility and cannot adapt to the quality differences and feature completeness variations of different fingerprint images. For example, for clear images with abundant details, details should dominate; while for blurry images or images with few details, the importance of tertiary features such as ridge shape should be increased. Fixed weight allocation mechanisms are difficult to achieve such dynamic adjustments, resulting in insufficient robustness in complex real-world scenarios.

[0005] Furthermore, when processing partial fingerprints, the design of existing technical processes often fails to effectively balance the contradiction between retrieval efficiency and matching accuracy. If a full comparison of fine features is performed directly in the entire database, the computational cost is huge and the response speed is slow. If only coarse-grained retrieval is performed, the final result may be affected by the excessively large candidate set or insufficient accuracy. At the same time, in the fine matching stage, effectively establishing the correlation between local features and the more macroscopic ridge structure to enhance the discriminative power and alignment accuracy of the matching is also a direction that needs to be improved in existing methods.

[0006] Therefore, how to design a partial fingerprint recognition method based on multi-level feature fusion, while ensuring large-scale retrieval efficiency, and improving matching accuracy under conditions of limited features and variable quality, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of this, the present invention provides a partial fingerprint recognition method based on multi-level feature fusion, which aims to solve the recognition problem caused by the small effective area and limited feature information of partial fingerprint images. Through a multi-level feature fusion matching process, it achieves efficient large-scale fingerprint database retrieval. On this basis, it utilizes the complementary advantages of different types of features to complete the accurate identification of partial fingerprints.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A partial fingerprint recognition method based on multi-level feature fusion includes the following steps: S1. Preprocess the input fingerprint image to obtain the preprocessed fingerprint image; S2. Based on the preprocessed fingerprint image, extract the first-level curvature features, the second-level minutiae features, and the third-level ridge shape features respectively; S3. The fingerprint template library is retrieved and classified using the first-level curvature features to obtain a set of candidate fingerprints; S4. Based on the candidate fingerprint set, perform minutiae matching and ridge shape matching using secondary minutiae features and tertiary ridge shape features to obtain minutiae matching scores and ridge shape matching scores. S5. Based on the minutiae matching score and ridge shape matching score, and combined with adaptive weight fusion, calculate the final matching score to obtain the fingerprint matching result.

[0010] Preferably, in step S2, extracting the first-order curvature feature includes: The preprocessed fingerprint image is divided into W×W sub-blocks, and the dominant orientation is estimated in each sub-block. ; Select the two neighboring blocks with the smallest angle to the center block and calculate the block curvature BC:

[0011] in, and These are the orientation angles of the two neighboring blocks, respectively. If the neighborhood direction variance of a block is greater than the threshold, it is marked as unreliable and curvature assignment is skipped.

[0012] Preferably, in step S2, extracting secondary minutiae features includes: Detecting ridge endpoints and bifurcation points as detail points Every detail Represented as:

[0013] in( , () are coordinates, For direction; For each detail point m, collect d sampling points around it. SIFT descriptors are calculated for each minutiae itself and for each of its sampled points; each minutiae... The complete descriptor is defined as:

[0014] in, This represents the SIFT descriptor, where d is the number of sampling points.

[0015] Preferably, in step S2, extracting the shape features of the tertiary ridge lines includes: The preprocessed fingerprint image is binarized and skeletonized to obtain the ridge skeleton. Measure the ridge width point by point on the ridge skeleton, compare the ridge width of each point with the width of its N neighboring points on both sides, and determine whether the point belongs to a concave or convex segment based on the width difference. If multiple consecutive points are identified as concave or convex, then the center point of that segment is extracted as a ridge feature point. :

[0016] in, As coordinates, In the direction of the ridge line, This refers to the ridge line morphology type.

[0017] Preferably, S3 includes: K-means clustering is performed on the first-level curvature feature vectors of all fingerprints in the template library to predetermine multiple categories; Calculate the first-order curvature feature vector of the fingerprint image to be identified; Based on the distance between the feature vector and the center of each category, it is assigned to the category with the closest distance, and all template fingerprints under that category are used as the candidate fingerprint set.

[0018] Preferably, in step S4, obtaining the minutiae matching score includes: For details in the template fingerprint image And query details in fingerprint images Local structures L containing the shape features of its neighboring details and neighboring ridges are constructed respectively. and L ; Using a dynamic programming strategy, by maximizing L and L The similarity between them is used to calculate the minutiae pairs. Local structural similarity score; The top N pairs of minutiae with the highest local structural similarity scores are selected as the initial matching pairs, and the final set of matching minutiae pairs is obtained by expanding them using a breadth-first search algorithm. Based on the final set of matched minutiae pairs, calculate the minutiae matching rate. and detail matching score If the detail matching rate Exceeding the preset threshold If the minutiae match is successful, the geometric transformation relationship between the two fingerprint images is estimated based on the matched minutiae pairs.

[0019] Preferably, the minutiae matching score Represented as:

[0020] in, For detail matching rate, To determine the number of successfully matched details, The total number of minutiae in the template fingerprint image. To query the total number of minutiae in a fingerprint image.

[0021] Preferably, in step S4, obtaining the ridge shape matching score includes: Based on the geometric transformation relationship between fingerprint images, the ridge shape feature points in the template fingerprint image are mapped to the coordinate system of the query fingerprint image; Within a preset radius of pixels at the mapped location, search for candidate ridge shape feature points in the query fingerprint image; Calculate the similarity score between the template ridge shape feature points and the query candidate points. If the score exceeds the preset threshold If the ridge shape feature points are matched successfully, a ridge shape matching score is obtained. .

[0022] Preferably, the ridge shape matching score Represented as:

[0023] in, This represents the number of successfully matched ridge feature points. Indicates the v-th pair of matching ridge points Similarity score, and These represent the total number of ridge-shaped feature points in the template and query fingerprint images, respectively. and The total number of ridge shape feature points in the estimated overlapping region of the template and query fingerprint images, respectively.

[0024] Preferably, S5 includes: Calculate the final matching score :

[0025] Where λ is the adaptive weighting factor. Scoring is done based on the details. Scoring is given based on the shape matching of the ridge line; Final match score Compared with the preset global decision threshold, if If the value is greater than the global decision threshold, then the query fingerprint image is determined to be a successful match with the current template fingerprint image.

[0026] As can be seen from the above technical solution, compared with the prior art, the technical solution of the present invention has the following beneficial effects: 1. This technical solution effectively balances the contradiction between efficiency and accuracy in some fingerprint recognition processes by adopting a multi-level feature fusion strategy. First, it uses the first-level curvature feature to quickly search and screen a large-scale template library, narrowing the comparison range. Then, it combines the second-level minutiae feature and the third-level ridge shape feature in the candidate set for fine matching, which improves the overall system efficiency while ensuring recognition accuracy.

[0027] 2. By introducing an adaptive weight fusion mechanism, the contribution of different feature levels can be dynamically adjusted according to the actual matching situation, thereby enhancing the adaptability of the method to fingerprint images of different quality. When there are enough minutiae, the focus is on the minutiae matching score, while when there are few minutiae or the image quality is poor, the weight of the ridge shape feature is increased, so that the system can still maintain stable recognition performance in some fingerprints, low quality or local occlusion scenarios.

[0028] 3. This method integrates macroscopic curvature features, local detail features, and microscopic ridge morphology features to construct a hierarchical feature representation system. This multi-level and multi-type feature complementarity mechanism can provide supplementary identification information through other features when a single feature is limited, thereby improving the overall discrimination capability of some fingerprint recognition. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0030] Figure 1 A flowchart of a partial fingerprint recognition method based on multi-level feature fusion provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the preprocessing process provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the spatial curvature calculation process provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a 3×3 block neighborhood provided in an embodiment of the present invention; Figure 5 A schematic diagram for estimating ridge width provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of a test segment for ridge shape feature extraction provided in an embodiment of the present invention; Figure 7 This is a schematic diagram illustrating the number of feature points based on the area of ​​the sensing region, provided in an embodiment of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] like Figure 1 As shown, this embodiment provides a partial fingerprint recognition method based on multi-level feature fusion, including the following steps: S1. Preprocess the input fingerprint image to obtain the preprocessed fingerprint image; S2. Based on the preprocessed fingerprint image, extract the first-level curvature features, the second-level minutiae features, and the third-level ridge shape features respectively; S3. The fingerprint template library is retrieved and classified using the first-level curvature features to obtain a set of candidate fingerprints; S4. Based on the candidate fingerprint set, perform minutiae matching and ridge shape matching using secondary minutiae features and tertiary ridge shape features to obtain minutiae matching scores and ridge shape matching scores. S5. Based on the minutiae matching score and ridge shape matching score, and combined with adaptive weight fusion, calculate the final matching score to obtain the fingerprint matching result.

[0033] This method effectively balances retrieval efficiency and matching accuracy in some fingerprint recognition scenarios through a multi-level feature screening and fusion strategy. It utilizes first-level features for rapid initial screening, combines second- and third-level features for fine comparison, and dynamically balances the contribution of different features in dealing with image quality differences through an adaptive weighting mechanism. This improves the overall recognition stability of the method in partial, low-quality, or occluded fingerprint scenarios.

[0034] Please provide a further detailed explanation of each step and related feature in the above method; In this embodiment, S1, the input fingerprint image is preprocessed to obtain a preprocessed fingerprint image; for example... Figure 2 As shown, it includes: Normalization: Set preset mean and variance for the input fingerprint image to standardize the image; Local orientation estimation: Using the normalized fingerprint image, the orientation field of the ridges is estimated; Local frequency estimation: Calculate the frequency image based on the estimated orientation field and the normalized fingerprint image; Region mask estimation: Divide each pixel in the normalized fingerprint image into two categories, recoverable and non-recoverable, and generate a region mask; Filtering: An enhanced fingerprint image is generated by applying multiple Gabor filters to the ridge and valley pixels of the normalized fingerprint image. These Gabor filters are specifically designed to match the frequency and orientation of the local ridges.

[0035] In this embodiment S2, based on the preprocessed fingerprint image, the first-level curvature features, the second-level minutiae features, and the third-level ridge shape features are extracted respectively. The extraction of first-order curvature features includes: like Figure 3 As shown, the curve includes two points: point A and point M. The distance from A to M, or the arc length along the curve between these two points, is represented by the symbol... The acute angle formed by the tangents to the curve at points A and B is specified as . The curvature of the curve at point A can be obtained using the following formula:

[0036] In this implementation, instead of directly estimating the pixel-level orientation, the fingerprint image is divided into sub-blocks of size W×W, and a dominant orientation is estimated for each sub-block. The value of W ranges from 16 to 32 pixels to achieve a balance between orientation consistency and detail preservation. The dominant orientation of each block is regarded as the tangent angle of the local ridge. All pixels in the block inherit the orientation of the block, thereby avoiding drastic fluctuations in pixel-level orientation due to noise and local texture interference. like Figure 4 As shown, within its 3×3 neighborhood, the curvature of the central block is defined by the discrete curvature formula as half the orientation difference between the two neighboring blocks in the oriented direction:

[0037] in, , To determine the direction of the neighboring block with the smallest angle to the direction of the central block, and to avoid confusion caused by multiple ridges in a large block, the direction variance is calculated on the four candidate direction axes of the central block. The axis with the smallest variance is selected as the representative direction, and the curvature is calculated accordingly. If the distribution of the neighborhood direction of a block is too discrete, the block is marked as unreliable and the curvature assignment is skipped.

[0038] Furthermore, the extraction of secondary minutiae features includes: Detail points are mainly divided into two categories: ridge endpoints and ridge bifurcation points. The former is the termination position of the ridge line, while the latter represents the node where a single path splits into two paths. Their description is defined by coordinates and direction angles.

[0039] Let m and This represents the detected details and sampling points. The descriptor for the detected details is defined as follows:

[0040] Where d represents the number of sampling points, It is a SIFT descriptor representing the sampling points around the detail point m. SIFT descriptors representing minutiae are particularly crucial in fingerprint verification, as they effectively characterize local texture information. To balance computational efficiency and verification accuracy, this embodiment sets the parameter d to 4. Compared to the original SIFT which can extract more than 2,000 key points, only about 50–200 descriptors are generated, reducing computational complexity while maintaining high discriminative power.

[0041] Furthermore, the extraction of tertiary ridge shape features includes: Ridge morphology can be extracted through concavity and convexity. Specifically, concave segments are characterized by narrower ridge widths, while convex segments are relatively wider. By comparing the width changes along the ridge direction, concave and convex areas can be effectively distinguished, thereby achieving the representation of ridge features. like Figure 5 As shown, the ridge width is estimated using the same binarized and skeletonized images of the fingerprint. In the binarized image, the ridge profile... Using a point in a skeletonized image The center, its length is And perpendicular to the direction of the local ridge line ( Figure 5 (as shown in a); if fewer than two boundary points are found at this location, then it is considered a ridge point. Unreliable, and skips the measurement of ridge width at that position; on the other hand, if at that position... If more than two boundary points are found, the procedure for identifying ridge boundaries is repeated at the same location in the enhanced image. Figure 5 (as shown in b) and The Euclidean distance between them is used to determine The ridge width W ( Figure 5 (as shown in c), therefore a binarized image is used to measure the ridge width; like Figure 6 As shown, on the refined ridge line, the width of the ridge line at each point is measured and compared with the width of its neighboring points to determine whether it belongs to a concave or convex segment. Let point x be the position of the narrow ridge line, and the difference in width between it and its two neighboring points N is used for discrimination. Figure 6 (a) shows that throughout the process, each neighboring point It is divided into three states to help identify the local morphology of the ridge line;

[0042] in, This represents the width of the ridge line at point x. Let represent the ridge width at neighboring point i, and let be the comparison threshold for the ridge width. According to the width of the ridge line and The comparison results show the state of ridge point i. It is specified as one of n, s, or w; To determine which points on a sparse ridge are concave or convex, the system performs a segmented test at each location. If a segment of ridge points is continuously classified as either fully concave or fully convex (…), the system will determine the ridge point. Figure 6 (b) shows that this segment is called a concave segment or a convex segment and is used to define ridge feature points. Conversely, if a single ridge point is identified as a concave or convex point, it will be considered a false detection and excluded due to noise or ambiguity. In a series of consecutive concave or convex points, the center point is the ridge feature point. :

[0043] in, Indicates the type of ridge shape (concave or convex). It is the direction of the ridgeline at that location. These are the coordinates of the center point. This feature representation is similar to the standard minutiae format used in fingerprint recognition; however, the directional range of the ridges is... The directional range of detail points is .

[0044] In this embodiment S3, the fingerprint template library is retrieved and classified using first-level curvature features to obtain a candidate fingerprint set; including: K-means clustering is performed on the first-level curvature feature vectors of all fingerprints in the template library to predetermine multiple categories; Calculate the first-order curvature feature vector of the fingerprint image to be identified; Based on the distance between the feature vector and the center of each category, it is assigned to the category with the closest distance, and all template fingerprints under that category are used as the candidate fingerprint set.

[0045] In this embodiment S4, based on the candidate fingerprint set, minutiae matching and ridge shape matching are performed using secondary minutiae features and tertiary ridge shape features to obtain minutiae matching scores and ridge shape matching scores. The detailed point matching score includes: Minute point matching finds the correspondence between a template and a query image by comparing their local neighborhoods, including based on a fixed number of nearby minute points or all neighbors within a certain range. However, when the number of minute points is insufficient, the discriminative power of these methods is limited and it is difficult to support reliable identification. Therefore, in this embodiment, minute points and their neighboring ridge features are jointly modeled to enhance the distinguishability of local structures. When the overlap area between two fingerprint images is small, even if part of the fingerprint image comes from the same finger, the nearest neighbor node of a minutiae may change. Therefore, within a specific range of the central minutiae, the proposed local structure includes all minutiae and ridge features. Assumption and These represent the minutiae in the template fingerprint image. And query details in fingerprint images Given two local structures that can work well together, a dynamic programming strategy can be used to find the optimal match to maximize... and Similarity score between them The formula for calculating the similarity score is as follows:

[0046] in, and Let represent the similarity obtained by comparing adjacent minutiae and adjacent ridge features, respectively. The similarity is determined using the following formula:

[0047] in, yes and The confidence score of the match between them and They are and The local neighborhood (details or ridge features) in the image, while and They represent and The total number of neighborhoods in the array.

[0048] In the template and query fingerprint images, the local structures based on minutiae are compared, and the top N pairs with the highest similarity scores are selected as initial matching pairs. Then, a breadth-first search algorithm is used to progressively find other corresponding minutiae from each initially matched minutiae pair, generating N different matching results. Finally, the matching result with the highest minutiae matching rate is selected. The minutiae matching rate is used to measure the degree of minutiae matching in the overlapping area of ​​two fingerprint images, and its calculation formula is as follows:

[0049] in, It is the number of matched details. It is a matching minutiae pair The equation-based comparison score, and These represent the total number of minutiae in the overlapping region of the template fingerprint image and the input fingerprint image, respectively. The overlapping region is estimated using matched minutiae pairs through a linear transformation. Exceeding the matching threshold If the initial matching minutiae pairs between the two images are correctly detected, these matching pairs are then used to make suggested ridge feature matching.

[0050] Further obtain detail matching scores :

[0051] in, For detail matching rate, The total number of minutiae in the template fingerprint image. To query the total number of minutiae in a fingerprint image.

[0052] Furthermore, the ridge shape matching score obtained in S4 includes: After completing the detail matching, it is necessary to further evaluate the correspondence between the ridge features in the overlapping area of ​​the template and the query image. Although detail matching has considered local ridge information, it is difficult to establish an accurate one-to-one correspondence by relying solely on the neighborhood of the detail points. By introducing non-local ridge feature comparison in the overlapping area, the accuracy of the overall matching can be effectively improved. set up Ridge features in the template fingerprint image, for The mapping representation on the input fingerprint image, taking into account transformation errors, may not contain any corresponding fingerprints in the input image. The perfectly matched ridge feature points, let for radius Ridge feature points within a pixel range, in order to determine and The correspondence is first represented as matching minutiae points. The set, The position and orientation in the template image are determined by the details of the template. Select the k nearest detail points to describe; Query details Similarly, it means Location and direction, and query details. The matching corresponds to the selected k template minutiae. If there are fewer than five matching minutiae pairs between two images, then all matching minutiae will be used for representation. and ,for and Next, using direction, radial angle, and relative distance to nearby matching detail point pairs, the following steps are determined. and Between :

[0053] in, yes The confidence score of the match, and yes and The nearest matching detail pair between them, if Exceeding the predetermined threshold Then the ridge features will be paired Save it in the candidate pair list and treat it as Candidate pairs, threshold The adjustment depends on the number k of nearby matching minutiae pairs;

[0054] The maximum and minimum values ​​are respectively represented by and It is stated that in order to reduce due to and Insufficient description leads to incorrect matches; as the k value decreases, the threshold... It will increase accordingly.

[0055] Finally, the ridge shape matching score is obtained. :

[0056] in, This represents the number of successfully matched ridge feature points. Indicates the v-th pair of matching ridge points Similarity score, and These represent the total number of ridge-shaped feature points in the template and query fingerprint images, respectively. and The total number of ridge shape feature points in the estimated overlapping region of the template and query fingerprint images, respectively.

[0057] The parameter settings for each stage of the feature extraction process are shown in Table 1: Table 1

[0058] In this embodiment, S5, based on the minutiae matching score and the ridge shape matching score, and combined with adaptive weight fusion, a final matching score is calculated to obtain the fingerprint matching determination result; including: Calculate the final matching score :

[0059] Where λ is the adaptive weighting factor. Scoring is done based on the details. Scoring is given based on the shape matching of the ridge line; Final match score Compared with the preset global decision threshold, if If the value is greater than the global decision threshold, then the query fingerprint image is determined to be a successful match with the current template fingerprint image.

[0060] Adaptive weighting factor Represented as:

[0061] in, The weights representing ridge features are used to assign specific relative importance to ridge features. =0.12), in this embodiment, the weight It embodies an adaptive allocation mechanism: when there are many available minutiae in the image, secondary features dominate the matching; when there are few minutiae or they are affected by noise, the weights of features such as tertiary ridges are increased accordingly, thereby compensating for insufficient local information, maintaining overall robustness, avoiding the rigidity of fixed weights, and being able to dynamically balance between different feature levels according to the sample quality, which helps to improve recognition performance under complex conditions such as low quality and local occlusion.

[0062] Furthermore, the experimental verification of this technical solution is based on publicly available benchmark datasets and a self-built database. A test environment is constructed by simulating partial fingerprint images of different sizes. FVC2002, FVC2004, and the self-built database are used. The original fingerprint image is cropped to generate three different sensory region areas: 9.3×9.3 mm², 7.7×7.7 mm², and 6.9×6.9 mm², corresponding to different image sizes, to simulate fingerprint information loss due to incomplete pressing or sensor limitations in real-world scenarios. Figure 7 As shown, as the area of ​​the sensing region decreases, the number of extractable details in the image decreases, which directly leads to the performance degradation of traditional matching methods that rely on details. This highlights the necessity of introducing multi-level complementary features in some fingerprint recognition methods.

[0063] The experimental design strictly followed common evaluation standards in the fingerprint recognition field, and large-scale real and fake matching tests were conducted on various databases. First, the template library was quickly searched and filtered using first-level curvature features, which significantly reduced the candidate set for subsequent fine matching. Then, within the candidate set, scores were calculated by combining second-level minutiae features and third-level ridge shape features through minutiae matching and ridge shape matching, respectively, and an adaptive weight fusion strategy was adopted to obtain the final matching result. To objectively evaluate the performance, the experiment selected the traditional minutiae matcher (CMM), minutiae cylindrical code (MCC), deep learning-based PoreNet, and the feature point detection method A-KAZE as comparison methods, and the equal error rate (EER) was used as the core evaluation metric.

[0064] Experimental results show that this technical solution achieves superior matching performance on partial fingerprint images of different sizes across multiple databases. The combined results are shown in Table 2 (fingerprint matching performance simulated on selected fingerprint datasets in the DB1 and DB3 databases of FVC2002), Table 3 (fingerprint matching performance simulated on selected fingerprint datasets in the DB1 and DB2 databases of FVC2004), and Table 4 (fingerprint matching performance simulated on selected fingerprint datasets in the database of this embodiment). Table 2

[0065] Table 3

[0066] Table 4

[0067] This method achieved the lowest EER under most test conditions, and its performance advantage was even more obvious in challenging scenarios with small fingerprint regions and a limited number of minutiae. In addition, this method can maintain a high true acceptance rate under different false acceptance rates, and it has good discrimination stability and robustness. The relevant experimental results jointly prove the effectiveness of the multi-level feature fusion and adaptive weighting mechanism in improving the accuracy and adaptability of partial fingerprint recognition in this embodiment.

[0068] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0069] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A partial fingerprint recognition method based on multi-level feature fusion, characterized in that, Includes the following steps: S1. Preprocess the input fingerprint image to obtain the preprocessed fingerprint image; S2. Based on the preprocessed fingerprint image, extract the first-level curvature features, the second-level minutiae features, and the third-level ridge shape features respectively; S3. The fingerprint template library is retrieved and classified using the first-level curvature features to obtain a set of candidate fingerprints; S4. Based on the candidate fingerprint set, perform minutiae matching and ridge shape matching using secondary minutiae features and tertiary ridge shape features to obtain minutiae matching scores and ridge shape matching scores. The detailed point matching score includes: For details in the template fingerprint image And query details in fingerprint images Local structures L containing the shape features of its neighboring details and neighboring ridges are constructed respectively. and L A dynamic programming strategy is adopted to maximize L. and L Similarity between them, calculate minutiae pairs The local structural similarity score is calculated; the top N pairs of minutiae with the highest local structural similarity scores are selected as initial matching pairs, and a breadth-first search algorithm is used to expand the set of final matching minutiae pairs; based on the final set of matching minutiae pairs, the minutiae matching rate is calculated. and detail matching score If the detail matching rate Exceeding the preset threshold If the minutiae match is successful, the geometric transformation relationship between the two fingerprint images is estimated based on the matched minutiae pairs. The ridge shape matching score includes: Based on the geometric transformation relationship between fingerprint images, ridge shape feature points in the template fingerprint image are mapped to the coordinate system of the query fingerprint image; within a preset radius pixel range of the mapped position, candidate ridge shape feature points are searched in the query fingerprint image; and the similarity score between the template ridge shape feature points and the query candidate points is calculated. If the score exceeds the preset threshold If the ridge shape feature points are matched successfully, a ridge shape matching score is obtained. ; S5. Based on the minutiae matching score and ridge shape matching score, and combined with adaptive weight fusion, calculate the final matching score to obtain the fingerprint matching determination result; including: Calculate the final matching score : Where λ is the adaptive weighting factor; Final match score Compared with the preset global decision threshold, if If the value is greater than the global decision threshold, then the query fingerprint image is determined to be a successful match with the current template fingerprint image.

2. The partial fingerprint recognition method based on multi-level feature fusion according to claim 1, characterized in that, In S2, extracting the first-order curvature features includes: The preprocessed fingerprint image is divided into W×W sub-blocks, and the dominant orientation is estimated in each sub-block. ; Select the two neighboring blocks with the smallest angle to the center block and calculate the block curvature BC: in, and These are the orientation angles of the two neighboring blocks, respectively. If the neighborhood direction variance of a block is greater than the threshold, it is marked as unreliable and curvature assignment is skipped.

3. The partial fingerprint recognition method based on multi-level feature fusion according to claim 1, characterized in that, In step S2, the extraction of secondary minutiae features includes: Detecting ridge endpoints and bifurcation points as detail points Every detail Represented as: in( , () are coordinates. For direction; For each detail point m, collect d sampling points around it. SIFT descriptors are calculated for each minutiae itself and for each of its sampled points; each minutiae... The complete descriptor is defined as: in, This represents the SIFT descriptor, where d is the number of sampling points.

4. The partial fingerprint recognition method based on multi-level feature fusion according to claim 1, characterized in that, In step S2, the extraction of the tertiary ridge shape features includes: The preprocessed fingerprint image is binarized and skeletonized to obtain the ridge skeleton. Measure the ridge width point by point on the ridge skeleton, compare the ridge width of each point with the width of its N neighboring points on both sides, and determine whether the point belongs to a concave or convex segment based on the width difference. If multiple consecutive points are identified as concave or convex, then the center point of that segment is extracted as a ridge feature point. : in, As coordinates, In the direction of the ridge line, This refers to the ridge line morphology type.

5. The partial fingerprint recognition method based on multi-level feature fusion according to claim 1, characterized in that, S3 includes: K-means clustering is performed on the first-level curvature feature vectors of all fingerprints in the template library to predetermine multiple categories; Calculate the first-order curvature feature vector of the fingerprint image to be identified; Based on the distance between the feature vector and the center of each category, it is assigned to the category with the closest distance, and all template fingerprints under that category are used as the candidate fingerprint set.

6. The partial fingerprint recognition method based on multi-level feature fusion according to claim 1, characterized in that, The minutiae matching score Represented as: in, For detail matching rate, To determine the number of successfully matched details, The total number of minutiae in the template fingerprint image. To query the total number of minutiae in a fingerprint image.

7. The partial fingerprint recognition method based on multi-level feature fusion according to claim 1, characterized in that, The ridge shape matching score Represented as: in, This represents the number of successfully matched ridge feature points. Indicates the v-th pair of matching ridge points Similarity score, and These represent the total number of ridge-shaped feature points in the template and query fingerprint images, respectively. and The total number of ridge shape feature points in the estimated overlapping region of the template and query fingerprint images, respectively.

Images