A method for latent fingerprint recognition and an apparatus thereof

Abstract

The application discloses a latent fingerprint identification method and device, comprising the following steps: inputting a latent fingerprint image into a trained generative adversarial network to obtain an enhanced latent fingerprint image; inputting the enhanced latent fingerprint image into a trained encoder for compression to obtain low-dimensional potential data; performing forward diffusion on the low-dimensional potential data, continuously adding Gaussian noise to the low-dimensional potential data to generate random noise, performing reverse diffusion on the random noise, and gradually restoring the random noise to the low-dimensional potential data; inputting the restored low-dimensional potential data into a trained decoder for decoding to obtain a repaired image; according to the characteristics of the fingerprint image, the diffusion model is improved, the fingerprint direction field constraint is added to solve the generation problem of irrelevant fingerprints, the fingerprint identification is facilitated, and the latent fingerprint image identification performance can be improved.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and specifically relates to a latent fingerprint recognition method and apparatus. Background Technology

[0002] Unlike rolled fingerprints collected with the cooperation of the person involved, latent fingerprints are generally of poor quality. At crime scenes, suspects may unintentionally leave fingerprints on surfaces. In such cases, the image quality is difficult to guarantee. Dust, water stains, blood, or other contaminants present at the scene often result in poor image quality, low contrast between the foreground and background, and an inability to clearly distinguish fingerprint texture from background dirt and noise. Furthermore, unlike conventional fingerprint images acquired through sensors, the physical or chemical methods used for latent fingerprint image acquisition cannot guarantee the integrity of the fingerprint image. Because fingerprints are left unintentionally, the condition of the finger (dry, wet, injured) and the surrounding environment (object type, surface cleanliness) are highly random. Combined with the influence of the uniformity of finger pressure, all these factors interfere with the formation of latent fingerprint images, leading to areas where imaging is impossible or where fingerprint texture is absent, resulting in incomplete latent fingerprints. The acquisition and imaging processes make latent fingerprint recognition technology a current challenge in this field.

[0003] While automatic fingerprint recognition systems are booming and widely used, current systems typically handle high-quality fingerprint images—those acquired by the fingerprint sensor with the subject's cooperation. These images have clear ridges, distinct valleys, and usually no background interference, showing only the fingerprint foreground. Low-quality latent fingerprint images, however, are characterized by poor texture quality, complex background noise, and incomplete fingerprint regions. These characteristics make it difficult for existing fingerprint recognition systems to handle latent fingerprints. During the recognition process, traditional image processing methods for high-quality fingerprint images are inadequate for complex, low-quality latent fingerprint images. In the feature extraction stage, the recognition system cannot accurately distinguish the fingerprint texture from the background pattern in latent fingerprint images, nor can it accurately locate and identify fingerprint detail features on the fingerprint ridges, ultimately leading to decreased latent fingerprint recognition accuracy and failure to meet performance requirements.

[0004] Therefore, a method for enhancing and repairing latent fingerprint images is provided to improve image quality and enhance recognition. Summary of the Invention

[0005] To address the aforementioned problems in the prior art, this invention provides a latent fingerprint recognition method and apparatus. The technical problem to be solved by this invention is achieved through the following technical solution:

[0006] In a first aspect, the present invention provides a latent fingerprint recognition method, comprising:

[0007] The latent fingerprint image is input into the generator for processing, and the latent fingerprint image is added as a constraint to the generator. The generator outputs the ridge texture image corresponding to the latent fingerprint image. The real ridge texture image corresponding to the latent fingerprint image is added as a constraint to the discriminator. The discriminator processes the ridge texture image corresponding to the latent fingerprint image and outputs the enhanced fingerprint image.

[0008] The enhanced fingerprint image is input into a trained encoder for compression to obtain low-dimensional latent data. The low-dimensional latent data is then forward-diffused, and Gaussian noise is continuously added to generate random noise. This random noise is then backward-diffused to gradually restore the low-dimensional latent data. The restored low-dimensional latent data is input into a trained decoder for decoding to obtain the repaired image. Furthermore, during the process of obtaining the repaired image from the enhanced fingerprint image, fingerprint orientation field constraints are added to address the problem of generating irrelevant fingerprints, thus enabling fingerprint recognition.

[0009] Secondly, the present invention also provides a latent fingerprint recognition device, comprising:

[0010] The fingerprint enhancement module is used to input the latent fingerprint image into the generator for processing, and to add the latent fingerprint image as a constraint to the generator. The generator outputs the ridge texture image corresponding to the latent fingerprint image. The real ridge texture image corresponding to the latent fingerprint image is added as a constraint to the discriminator. The discriminator processes the ridge texture image corresponding to the latent fingerprint image and outputs the enhanced fingerprint image.

[0011] The latent fingerprint image restoration module is used to input the enhanced latent fingerprint image into a trained encoder for compression to obtain low-dimensional latent data. The low-dimensional latent data is then forward-diffused, and Gaussian noise is continuously added to generate random noise. This random noise is then backward-diffused to gradually restore the random noise to low-dimensional latent data. The restored low-dimensional latent data is then input into a trained decoder for decoding to obtain the restored image. Furthermore, during the process of obtaining the restored image from the enhanced fingerprint image, fingerprint orientation field constraints are added to solve the problem of generating irrelevant fingerprints, thus enabling fingerprint recognition.

[0012] The beneficial effects of this invention are:

[0013] This invention provides a latent fingerprint recognition method and apparatus, proposing a combined approach of latent fingerprint enhancement and restoration. The enhanced image undergoes restoration processing to further improve fingerprint image recognition performance. A targeted improvement to a deep network image restoration method is applied to latent fingerprint restoration. Fingerprint orientation field constraints are added to the model to reduce the generation of irrelevant fingerprints and avoid generating large erroneous ridges that could affect subsequent fingerprint image recognition and matching. Compared to traditional image restoration methods and deep learning generation methods such as GANs, this method produces excellent restored fingerprint images, facilitating further recognition applications. For unknown fingerprint datasets, Stable Diffusion demonstrates excellent generalization ability, showing good restoration results across different fingerprint datasets. For texture images, both the thinned and acquired images exhibit excellent performance.

[0014] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0015] Figure 1 This is a flowchart of a latent fingerprint recognition method provided in an embodiment of the present invention;

[0016] Figure 2 This is a schematic diagram of a latent fingerprint recognition method provided in an embodiment of the present invention;

[0017] Figure 3 This is a schematic diagram of a latent fingerprint enhancement algorithm model provided in an embodiment of the present invention;

[0018] Figure 4 This is a schematic diagram of latent fingerprint image restoration provided in an embodiment of the present invention;

[0019] Figure 5 This is a schematic diagram of latent fingerprint image restoration provided in an embodiment of the present invention;

[0020] Figure 6 This is a schematic diagram showing a comparison before and after repair provided in an embodiment of the present invention;

[0021] Figure 7 This is a schematic diagram of CMC curves for fingerprint repair of different missing sizes provided in an embodiment of the present invention;

[0022] Figure 8 This is a schematic diagram of a NIST SD27 latent fingerprint provided in an embodiment of the present invention;

[0023] Figure 9 This is a schematic diagram of latent fingerprints and enhancement results provided in an embodiment of the present invention. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0025] In related technologies, latent fingerprint recognition is mainly performed from the perspective of fingerprint image enhancement, including background removal and image ridge enhancement. However, current enhancement algorithms have difficulty extracting complete ridge images from the background when there is strong noise interference, and most ridges are not continuous and complete. Therefore, this invention proposes to enhance incomplete fingerprint images first and then repair them.

[0026] Traditional image inpainting algorithms are mainly divided into two categories: diffusion-based inpainting methods and patch-based inpainting methods. Diffusion-based image inpainting gradually diffuses the pixel information of the existing portion of the incomplete image outwards, smoothly propagating the image content from the existing area to the missing hole. Patch-based image inpainting assumes that the remaining areas in the image contain the content and structure of the damaged area. The algorithm finds the patch with the greatest similarity to the missing area from the existing area, copies the pixel information, and uses it to fill in the missing structure. Milind proposed an enhanced KNN-SVD method to inpaint a very small portion of the acquired fingerprint image.

[0027] Deep learning-based methods: In 2016, Pathak proposed a content encoder network that uses context-based pixel prediction to drive the network to learn image inpainting, enabling it to understand image semantics to a certain extent and laying the foundation for future research. Inspired by this, Satoshi proposed a strategy based on local and global consistency to address the inconsistency in context-encoded inpainting regions. Wang used branch networks with different sampling rates to extract features at different levels, overcoming the problem that a single network can only extract features at the same level. Yan added shift connections to connect the blank search of the encoding layer to the decoding layer, generating a fine texture structure through information sharing. Zeng used a pyramid structure to understand high-level semantic information of the image, ensuring visual and semantic consistency of the restored region. Zhuang proposed four CNN models for enhancing and inpainting ODF images, using explicit fingerprint (ODF) sensor images as the research object.

[0028] The deep learning-based methods described above can be broadly categorized into Context-Encoder, U-Net, and adversarial model types based on their core architectures. Context-Encoder is an unsupervised visual feature model. Its content encoder utilizes contextual information from damaged regions and known content, combined with global image features, to generate a structure that matches existing content. U-Net-type structures, improved from FCN networks, were first applied to image segmentation. U-Net constructs cross-layer feature connections at the same resolution in encoding and decoding, fusing low-level feature information from the encoding stage with high-level abstract features from the decoding stage, compensating for information loss caused by downsampling. Generative Adversarial Networks (GANs), by introducing game theory mechanisms and adversarial learning strategies, free the network's generative capabilities from direct limitations imposed by training data, enabling the generation of realistic images.

[0029] Diffusion models are generative models used across various deep learning domains. Currently, they are primarily used for image and audio generation, and some literature has already used diffusion deep models for image inpainting. Lugmayr et al. proposed a method for performing high-quality, diverse image inpainting, even for extreme masks, using a pre-trained unconditional denoising diffusion probability model as prior generation. W Li et al. proposed a method for inpainting large missing regions while preserving image texture and structure, utilizing Markov random fields for region inpainting. All of the above literature focuses on inpainting natural images, but there is currently no literature specifically addressing fingerprint image inpainting.

[0030] Since the restoration of incomplete fingerprint images is still a relatively new field, there are not many existing technologies. As for the enhanced KNN-SVD method mentioned above, its purpose is to restore fingerprint images of fingers with minor cuts or burns. The incomplete area is small, which has little impact on fingerprint recognition, and the effect is relatively average.

[0031] Traditional image restoration methods, such as diffusion and patching, essentially aim to find compensation information from known image areas to replace missing content. The difference lies in the search scope: diffusion only searches around the boundaries of the damaged content, extending the search to undamaged areas, while patching expands the search to the entire image, copying similar content from undamaged regions.

[0032] In view of this, and considering that existing technologies are limited to enhancing latent fingerprints to improve recognition performance, the present invention provides a latent fingerprint recognition method that performs repair processing after latent fingerprint enhancement, further increasing the number of detail points in the fingerprint image to improve recognition performance and thus improve fingerprint recognition accuracy.

[0033] Please see Figure 1 and Figure 2 , Figure 1 This is a flowchart of a latent fingerprint recognition method provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of a latent fingerprint recognition method provided by an embodiment of the present invention. The latent fingerprint recognition method provided by the present invention includes:

[0034] S101. Input the latent fingerprint image into the generator for processing, and add the latent fingerprint image as a constraint to the generator. The generator outputs the ridge texture image corresponding to the latent fingerprint image. Add the real ridge texture image corresponding to the latent fingerprint image as a constraint to the discriminator. The discriminator processes the ridge texture image corresponding to the latent fingerprint image and outputs the enhanced fingerprint image.

[0035] For details, please continue to see Figure 2 In this embodiment, the target is latent fingerprint images. Due to the noisy nature of latent fingerprint images, in order to perform better restoration, the latent fingerprint images are first enhanced to extract clearer fingerprint texture features, and then image restoration is performed on this basis.

[0036] In this embodiment, the powerful learning and representation capabilities of generative adversarial networks (GANs) in fingerprint generation tasks are leveraged to enhance images, thereby constructing the required fingerprint ridge enhancement network structure. To make the generative model more adaptable to the fingerprint ridge extraction task, constraints are added to the generator's training and learning. The latent fingerprint image is incorporated as a constraint into the input, and random noise, originally used as input, is discarded to avoid additional data interference in fingerprint ridge generation. The generator's output is a clear ridge texture image corresponding to the latent fingerprint image. The discriminator, in addition to improving the realism of fingerprint ridge generation, also constrains the ridge texture to the corresponding latent fingerprint. Thus, by adding latent fingerprint constraints, the GAN's ability to generate random inputs is associated with the original latent fingerprint image, constraining the task to the generation of ridge texture structures in the latent fingerprint image. By discarding random noise input, the generator is functionally transformed into an end-to-end fingerprint ridge extractor, completing the fingerprint enhancement task.

[0037] Please see Figure 3 , Figure 3This is a schematic diagram of a latent fingerprint enhancement algorithm model provided in an embodiment of the present invention. The latent fingerprint enhancement algorithm model includes a generator and a discriminator. The latent fingerprint image is input into the generator to generate images with different resolutions. The discriminator judges the images with different resolutions to enhance the latent fingerprint image and outputs the enhanced latent fingerprint image.

[0038] S102. The enhanced latent fingerprint image is input into the trained encoder for compression to obtain low-dimensional latent data. The low-dimensional latent data is forward diffused, and Gaussian noise is continuously added to the low-dimensional latent data to generate random noise. The random noise is then backward diffused to gradually restore the random noise to low-dimensional latent data. The restored low-dimensional latent data is input into the trained decoder for decoding to obtain the repaired image. In this embodiment, in the process of obtaining the repaired image from the enhanced fingerprint image, the diffusion model is improved according to the characteristics of the fingerprint image, and fingerprint orientation field constraints are added to solve the problem of generating irrelevant fingerprints for fingerprint recognition.

[0039] Specifically, please see Figure 4 , Figure 4 This is a schematic diagram of latent fingerprint image restoration provided in an embodiment of the present invention. In this embodiment, the Stable Diffusion model is used for fingerprint image restoration. Stable Diffusion belongs to a class of deep learning models called diffusion models. Diffusion models are likelihood-based models that have recently been proven to generate high-quality images. Most of the diffusion models currently used come from the 2020 work DDPM: Denoising Diffusion Probabilistic Models. It can progressively denoise random Gaussian noise to obtain samples of interest, such as unconditional image synthesis, inpainting, super-resolution, class-conditional image generation, text-to-image generation, and layout-to-image generation. This embodiment mainly uses the image restoration function.

[0040] The diffusion model can be divided into two processes: forward diffusion and reverse diffusion. The forward diffusion process adds noise to the image, while the reverse diffusion process removes noise from the image.

[0041] The forward diffusion process involves continuously adding Gaussian noise to the original image x0 to generate random noise x. tThe process involves continuously adding Gaussian noise to low-dimensional latent data to generate random noise x. t ; where x t-1 To x t The process can be represented as:

[0042]

[0043] Where, α i denoted as a very small hyperparameter, i represents the index of the hyperparameter, t represents the step size, ∈~N(0,1) represents Gaussian noise, and N represents the normal distribution, i.e., the symbol for the Gaussian distribution.

[0044] From the above formula, we can derive x0 to x t The process is expressed as:

[0045]

[0046] Where x0 represents low-dimensional latent data, x t Indicates random noise.

[0047] The forward diffusion process transforms low-dimensional latent data into random noise, while the reverse process transforms random noise x by predicting Gaussian noise ∈. t The data is gradually reduced to low-dimensional latent data x0; the backdiffusion process is represented as follows:

[0048]

[0049] Where, x t-1 Represents random noise x t The result after removing noise once, x0 is the original image without noise, x1 is the image after the first addition of noise, x t For the image after adding noise for the tth time, ∈θ(x) t (t) represents a noise estimation model with training parameters θ, used to estimate the true Gaussian noise ∈ σ. t z represents the error between the predicted Gaussian noise and the actual Gaussian noise, σ t =1-α t , z~N(0,1).

[0050] The key to the back diffusion process is training the noise estimation model. θ Used to estimate Gaussian noise ∈, noise estimation model ∈ θ The loss function can be the MSE error, and its expression is:

[0051]

[0052] This embodiment employs a modified version of the diffusion model, Stable Diffusion. Stable Diffusion typically uses a U-Net neural network for training, which consists of an encoder and a decoder composed of residual modules. The encoder compresses the image, while the decoder decodes the low-resolution image into a high-resolution image. To prevent U-Net from losing important information during downsampling, a connection is added between the encoder's downsampling and the decoder's upsampling. The original diffusion model iteratively provides the full-size image to U-Net during the diffusion process to obtain the final result. When the total number of diffusion steps T and the image size are large, this pure diffusion model becomes very slow. Stable Diffusion improves upon this by encoding the image into latent data and then performing forward and backward diffusion processes in the latent space, thus overcoming the model's slowdown.

[0053] In this embodiment, a trained diffusion model is used to continuously predict noise in the image and denoise the low-dimensional latent data. After a certain number of steps, the noise-removed low-dimensional latent data is obtained. The latent data can then be decoded back into the image using a trained decoder E.

[0054] In an optional embodiment of the present invention, please refer to Figure 5 , Figure 5 This is a schematic diagram of latent fingerprint image restoration provided in an embodiment of the present invention. In the process of obtaining the restored image from the enhanced latent fingerprint image, a direction field constraint is added to control the range of the generated ridge direction within the direction field; wherein, the expression for calculating the ridge direction is:

[0055]

[0056] Among them, G x This indicates that the Sobel operator S x The calculated gradient in the x-direction, G y This indicates that the Sobel operator S y The calculated gradient in the y-direction, where I represents the enhanced fingerprint image, * represents the convolution operation, and G... xy G xx and G yy J represents intermediate calculation variables respectively. w This represents a matrix of all ones related to the magnitude of the direction field estimation, atan2(·) represents the arctangent calculation, and Θ I Indicates the direction of the output ridgeline.

[0057] The predicted angle at (x,y) can be represented as an N-dimensional vector. Where the i-th element pori (i) represents the probability of the ridge direction at that location, where The average ridge direction vector can be calculated as:

[0058]

[0059]

[0060]

[0061]

[0062] The loss for fingerprint orientation field constraints includes cross-entropy loss and orientation consistency loss; among which,

[0063] Cross-entropy loss function L CE The expression is:

[0064]

[0065] Directional consistency loss function L Coh The expression is:

[0066]

[0067] Here, ROI represents the region of interest. The labels of the label map tensor of the region of interest are summed at the corresponding positions in each channel. Positions with a value greater than 0 are set to 1, and the rest are set to 0, resulting in a binary label tensor. |ROI| represents the sum of this binary label tensor. This represents the summation of all elements in the matrix after multiplying the Coh and the label values ​​at the corresponding positions in the region of interest; λ + and λ - These are the weights of positive and negative samples. and p * (i|(x,y) represent the probabilities at (x,y) in the label map and the prediction map, respectively; J3 represents a 3*3 matrix of all ones; the label map refers to the manually labeled detail map, including the detail point score map, the x and y direction probability map, and the detail point angle distribution map.

[0068] L Total =λLOSS+L CE +L Coh ;

[0069] Where λ is the weighting coefficient.

[0070] Specifically, in this embodiment, an autoencoder is first trained to learn how to compress image data into a low-dimensional representation. Using the trained encoder E, a full-size image can be encoded into low-dimensional latent data (compressed data). Diffusion models alone are not very suitable for fingerprint images; therefore, this embodiment adds a fingerprint orientation field constraint. The fingerprint orientation field is an inherent property of fingerprint images, reflecting the center of the fingerprint image field, the trend field of circular ridges, the trend field of oblique-angle straight ridges, the trend field of horizontal ridges, and the mixed trend field. The fingerprint orientation field reflects the average orientation of ridges within a small fingerprint region, providing a rough description of the fingerprint ridges. As prior knowledge in the fingerprint domain, the orientation field constrains the fingerprint ridge generation process, limiting the range of generated ridges to the area surrounding the orientation field. The calculation of the fingerprint orientation field uses a gradient-based method and is transformed into a shallow neural network by replacing gradient calculation and windowing with convolution operations.

[0071] Adding the orientation field loss to the total loss can effectively solve the problem of generating unrelated fingerprints.

[0072] In this embodiment, since the input image of the present invention is generally noise-free, it is necessary to add noise to the low-dimensional latent data after encoder compression so that the diffusion model can generate a new image through denoising. In fingerprint image restoration, it is only necessary to add noise to the defective area and remove the noise through a backdiffusion process to restore the defective image.

[0073] In summary, this invention proposes a combined approach of latent fingerprint enhancement and restoration. The enhanced image undergoes restoration processing to further improve fingerprint image recognition performance. A targeted improvement to deep network image restoration methods is applied to latent fingerprint restoration. Fingerprint orientation field constraints are added to the model to reduce the generation of irrelevant fingerprints and avoid generating large erroneous ridges that could affect subsequent fingerprint image recognition and matching. Compared to traditional image restoration methods and deep learning generation methods such as GANs, this approach yields excellent restored fingerprint images, facilitating further recognition applications. For unknown fingerprint datasets, Stable Diffusion demonstrates excellent generalization ability, showing good restoration results across different fingerprint datasets. For texture images, both the thinned and acquired images exhibit excellent performance.

[0074] In an optional embodiment of the present invention, the effectiveness of the latent fingerprint recognition method provided in the above embodiment is verified by simulation experiments, specifically as follows:

[0075] In practical applications, when faced with the challenging task of filling in important missing parts of an image, most image inpainting methods use specific masks and fail to generalize well across different mask types. Compared to other methods, the diffusion model provided in this invention has better generalization ability because it is a probabilistic generative model that does not require mask-specific training and produces high-fidelity output.

[0076] Regarding fingerprint image restoration, this invention repairs fingerprint images in three different forms (normal acquisition, texture, and thinning). Experimental results demonstrate a visually satisfactory effect. Furthermore, a comparison of SSIM, PSNR, and LPIPS between the restored and damaged images, as shown in Table 1, reveals that the method employed in this invention significantly improves fingerprint image restoration. Further matching experiments confirm that the restoration process enhances fingerprint recognition rates. Moreover, a comparison of the NFIQ2.0 fingerprint image quality evaluation index before and after restoration further demonstrates the performance of this method.

[0077] Table 1 Comparison Chart

[0078]

[0079] On the one hand, please see Figure 6 , Figure 6 This is a schematic diagram showing the comparison before and after repair provided in an embodiment of the present invention. The effectiveness of the latent fingerprint recognition method provided by the present invention is verified by the repair effect diagram.

[0080] On the other hand, please see Figure 7 , Figure 7 This is a schematic diagram of CMC curves for fingerprint repair with different missing sizes provided in an embodiment of the present invention, verifying the effectiveness of the latent fingerprint recognition method provided by the present invention.

[0081] In an optional embodiment of the present invention, please refer to Figure 8 and Figure 9 , Figure 8 This is a schematic diagram of a NIST SD27 latent fingerprint provided in an embodiment of the present invention. Figure 9 This is a schematic diagram of latent fingerprints and enhancement results provided in an embodiment of the present invention. The present invention uses the general NIST SD27 latent fingerprint database set for experiments to evaluate our algorithm. The NIST SD27 dataset is collected from crime scenes, and the fingerprint images have low imaging quality and varying degrees of distortion. It is mainly used for fingerprint matching in forensic groups and criminal investigations. Figure 8 The image shown is the acquired latent fingerprint image. Figure 9 The image shown is a fingerprint image enhanced and repaired using the method provided by this invention, which clearly displays the fingerprint ridges.

[0082] Based on the same inventive concept, the present invention also provides a latent fingerprint recognition device for implementing the latent fingerprint recognition method provided in the above embodiments of the present invention. Specific embodiments of the method are described above and will not be repeated here. The device includes:

[0083] The fingerprint enhancement module is used to input the latent fingerprint image into the generator for processing, and to add the latent fingerprint image as a constraint to the generator. The generator outputs the ridge texture image corresponding to the latent fingerprint image. The real ridge texture image corresponding to the latent fingerprint image is added as a constraint to the discriminator. The discriminator processes the ridge texture image corresponding to the latent fingerprint image and outputs the enhanced fingerprint image.

[0084] The latent fingerprint image restoration module is used to input the enhanced latent fingerprint image into a trained encoder for compression to obtain low-dimensional latent data. The low-dimensional latent data is then forward-diffused, and Gaussian noise is continuously added to generate random noise. This random noise is then backward-diffused to gradually restore the random noise to low-dimensional latent data. The restored low-dimensional latent data is then input into a trained decoder for decoding to obtain the restored image. Furthermore, during the process of obtaining the restored image from the enhanced fingerprint image, fingerprint orientation field constraints are added to solve the problem of generating irrelevant fingerprints, thus enabling fingerprint recognition.

[0085] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or device comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device comprising said element. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect. The orientations or positional relationships indicated by terms such as "upper," "lower," "left," and "right" are based on the orientations or positional relationships shown in the accompanying drawings and are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0086] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0087] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for latent fingerprint recognition, characterized in that, include: The latent fingerprint image is input into the generator for processing, and the latent fingerprint image is added as a constraint to the generator. The generator outputs the ridge texture image corresponding to the latent fingerprint image. The real ridge texture image corresponding to the latent fingerprint image is added as a constraint to the discriminator. The discriminator processes the ridge texture image corresponding to the latent fingerprint image and outputs an enhanced fingerprint image. The enhanced fingerprint image is input into a trained encoder for compression to obtain low-dimensional latent data. The low-dimensional latent data is forward diffused, and Gaussian noise is continuously added to the low-dimensional latent data to generate random noise. The random noise is then reverse diffused to gradually restore the random noise to low-dimensional latent data. The restored low-dimensional latent data is input into a trained decoder for decoding to obtain the repaired image. Furthermore, during the process of obtaining the repaired image based on the enhanced fingerprint image, fingerprint orientation field constraints are added to solve the problem of generating irrelevant fingerprints, so as to perform latent fingerprint recognition.

2. The latent fingerprint recognition method according to claim 1, characterized in that, The forward diffusion process is represented as follows: Where x0 represents the low-dimensional potential data, x t This represents the random noise. α i Let represent the hyperparameter, i represent the index of the hyperparameter, t represent the step size, and ∈~N(0,1) represent Gaussian noise.

3. The latent fingerprint recognition method according to claim 2, characterized in that, The reverse diffusion process is represented as follows: Where, x t-1 Represents random noise x t The result after removing noise once represents the noise estimation model under the training parameters θ, used to estimate the true Gaussian noise ∈, σ. t z represents the error between the predicted Gaussian noise and the actual Gaussian noise, σ t =1-α t , z~N(0,1).

4. The latent fingerprint recognition method according to claim 3, characterized in that, The noise estimation model ∈ θ The loss function Loss is expressed as:

5. The latent fingerprint recognition method according to claim 1, characterized in that, In the process of obtaining the repaired image based on the enhanced latent fingerprint image, a fingerprint orientation field constraint is added to control the range of the generated ridge direction within the orientation field; wherein, the expression for calculating the ridge direction is: Among them, G x This indicates that the Sobel operator S x The calculated gradient in the x-direction, G y This indicates that the Sobel operator S y The calculated gradient in the y-direction, where I represents the enhanced fingerprint image, * represents the convolution operation, and G... xy G xx and G yy J represents intermediate calculation variables respectively. w This represents a matrix of all ones related to the magnitude of the direction field estimation, atan2(·) represents the arctangent calculation, and Θ I Indicates the direction of the output ridgeline; The predicted angle at (x,y) can be represented as an N-dimensional vector. Where the i-th element p ori (i) represents the probability of the ridge direction at that location; where The average ridge direction vector can be calculated as:

6. The latent fingerprint recognition method according to claim 5, characterized in that, The loss for the fingerprint orientation field constraint includes cross-entropy loss and orientation consistency loss; wherein, Cross-entropy loss function L CE The expression is: Directional consistency loss function L Coh The expression is: Here, ROI represents the region of interest. The labels of the ROI label map tensor are summed at the corresponding positions in each channel, with values ​​greater than 0 set to 1 and all other values ​​set to 0, resulting in a binary label tensor. |ROI| represents the sum of this binary label tensor. This represents the summation of all elements in the matrix after multiplying the Coh and the label values ​​at the corresponding positions in the region of interest; λ + and λ - These are the weights of positive and negative samples. and p * (i|(x,y) represent the probabilities at (x,y) in the label map and the prediction map, respectively; J3 represents a 3*3 matrix of all 1s; the label map refers to the manually labeled detail map, including the detail point score map, the x and y direction probability map, and the detail point angle distribution map; The total loss function is expressed as follows: L Total =λLOSS+L CE +L Coh ; Where λ is the weighting coefficient.

7. The latent fingerprint recognition method according to claim 1, characterized in that, In the process of obtaining the repaired image from the enhanced latent fingerprint image, the Stable Diffusion model is used, and the neural network used in the training process of the Stable Diffusion model is U-Net.

8. A latent fingerprint recognition device, characterized in that, include: The fingerprint enhancement module is used to input a latent fingerprint image into a generator for processing, and to add the latent fingerprint image as a constraint to the generator, which outputs a ridge texture image corresponding to the latent fingerprint image; and to add the real ridge texture image corresponding to the latent fingerprint image as a constraint to a discriminator, which processes the ridge texture image corresponding to the latent fingerprint image and outputs an enhanced fingerprint image. The latent fingerprint image restoration module is used to input the enhanced latent fingerprint image into a trained encoder for compression to obtain low-dimensional latent data, forward diffuse the low-dimensional latent data, continuously add Gaussian noise to the low-dimensional latent data to generate random noise, and reverse diffuse the random noise to gradually restore the random noise to low-dimensional latent data. The restored low-dimensional latent data is input into a trained decoder for decoding to obtain the repaired image. Furthermore, during the process of obtaining the repaired image based on the enhanced fingerprint image, fingerprint orientation field constraints are added to solve the problem of generating irrelevant fingerprints, so as to perform latent fingerprint recognition.

Images