Methods for processing proof images, related computer devices, and related computer programs

Abstract

The present invention relates to a method for processing test images showing multiple potential fingerprints, comprising the following steps: - Step (E102) involves generating a semantic segmentation map using a neural network, which includes a score indicating the probability of the presence of a latent fingerprint passing through each block of a test image. - A step (E104) to generate a binary map for each block of a plurality of blocks to estimate whether or not the block represents a latent fingerprint portion, - Step (E106) of detecting multiple instances of latent fingerprints from a binary map, - For each instance of multiple instances, the total score is obtained by summing the scores associated with the blocks among the multiple blocks belonging to the instance (E108), - The step of obtaining a limited number of instances by restricting the number of instances to each instance whose total score exceeds a threshold (E110) Regarding methods including

Description

Technical Field

[0001] The present invention relates to a method for processing a proof image showing a plurality of latent traces. Further, the present invention relates to a computer device and a related computer program.

[0002] Methods for identifying or authenticating an individual using the papillary traces (e.g., fingerprints and / or palm prints and / or footprints) of an individual are known, and these methods include the following steps. Conventionally, a test image showing a test fingerprint of an individual is obtained, and this image is compared with a reference image showing a reference fingerprint to verify a match between the test fingerprint and the reference fingerprint. If such a match is recognized, the individual is considered to be previously registered.

[0003] In the context of a police investigation, a photograph of one or more papillary traces left on a medium by the addition of an area of skin showing papillary ridges can be used as a test image. This photograph is typically obtained by a general device, i.e., a device not specifically designed to acquire papillary traces.

[0004] The papillary traces seen in this type of image are commonly referred to as "latent" fingerprints and can be likened to papillary fingerprints.

[0005] When comparing a reference image showing a reference fingerprint with a test image showing several latent traces, the performance of the identification and authentication process (typically, the false negative rate and / or the false positive rate) decreases.

[0006] To improve this, it is known to manually separate the latent traces from the test image, i.e., to instantiate the latent traces from the test image, and the latent traces can be individually processed by the identification and authentication process.

[0007] This separation is performed manually by an expert and is not very efficient for a large number of test images.

[0008] The paper "Automatic Latent Fingerprint Segmentation" by Dinh-Luan Nguyen, Kai Cao, and Anil K. Jain (arXiv:1804.09650v2) describes the use of neural networks to perform instance segmentation of fingerprints and the possibility of processing test images showing multiple latent fingerprints and separating the latent fingerprints from these test images. However, instance segmentation neural networks are complex to implement and expensive, particularly in terms of computation time and memory space.

[0009] Furthermore, test images may show several types of nipple prints, typically one or more latent fingerprints, and one or more latent palm prints.

[0010] To overcome these drawbacks, in a first embodiment, the present invention relates to a method for processing proof images showing multiple latent traces, which is carried out by a computer device. - A step of generating a semantic segmentation map from a test image by a neural network of semantic segmentation maps, for each block that forms part of multiple blocks in the test image, which includes a score associated with the block indicating the probability of presence of a latent fingerprint passing through the block in the test image. - A step of generating a binary map from a semantic segmentation map, wherein the binary map estimates whether each block of multiple blocks in the evidence image represents part of a latent fingerprint. - A step to detect multiple instances of latent traces from a binary map, - A step of obtaining a total score from a semantic segmentation map for each instance of multiple instances, wherein the total score is obtained by summing the scores associated with each block among the multiple blocks of the proof image belonging to that instance, - A step of obtaining a limited number of instances by restricting the number of instances to each instance whose total score exceeds a predetermined threshold. We propose a method that includes this.

[0011] The advantageous, unrestricted characteristics are as follows: - The total instance score is the average of the scores associated with a block among multiple blocks in the test image that represent part of the potential fingerprint of this detected instance. - The step of detecting multiple latent fingerprint instances includes a substep of applying watershed segmentation to the binary map using a sliding window. - The step of detecting multiple latent fingerprint instances includes a substep of applying linked component analysis to another binary map obtained from the binary map. - The step of detecting multiple latent fingerprint instances further includes a substep of obtaining other binary maps by eroding a binary map with an erosion window having a predetermined region, and removing blocks from a proof image that show a portion of the latent fingerprint having a region smaller than the predetermined region. - The method includes a step of reducing the number of instances between the step of detecting multiple instances of a latent trace and the step of obtaining a total score for each instance of the multiple instances. - The step of reducing multiple instances includes a substep for each instance of the multiple instances, which defines a detection box, the detection box being the smallest rectangular box that completely contains this instance. - The step of reducing multiple instances includes a substep of removing an instance from multiple instances if its discovery box is included in the discovery box of another instance from those multiple instances. - The step of reducing multiple instances includes a substep of merging the first instance and the second instance from the multiple instances if the intersection of the detection box of the first instance and the detection box of the second instance exceeds a predetermined ratio of the combined detection box of the first instance and the detection box of the second instance. - The method further includes an identification step of individually comparing at least one latent fingerprint instance from a limited number of instances with a reference fingerprint and verifying a match between each latent fingerprint from at least one latent fingerprint instance and the reference fingerprint.

[0012] According to a second aspect, the present invention proposes a computer program that includes a code instruction that performs the method of processing a proof image as described above, when the code instruction is executed by a computer.

[0013] This program may use any programming language and may be in the form of source code, object code, or intermediate code between source code and object code, such as a partially compiled form, or any other desired form.

[0014] According to a third aspect, the present invention provides a computer-readable storage medium for storing instructions that can be executed by a computer performing the proof image processing method as described above.

[0015] Computer-readable storage media are typically tangible or persistent media and may include storage media (e.g., hard disk drives, magnetic tape devices, or semiconductor memory devices).

[0016] According to a fourth aspect, the present invention relates to a computer device for processing proof images showing a plurality of latent traces, - A semantic segmentation map is generated from the test image using a neural network. The semantic segmentation map includes a score associated with each block that forms part of multiple blocks in the test image, indicating the probability of the latent fingerprint passing through the block in the test image. - A binary map is generated from the semantic segmentation map, and the binary map estimates whether each block in the evidence image represents part of a potential fingerprint. -Detect multiple instances of latent fingerprints from a binary map, - For each instance of multiple instances, obtain the total score from the semantic segmentation map, and then sum the scores associated with the blocks among the multiple blocks of the test image belonging to this instance to obtain the total score. - Obtain a limited number of instances by restricting the number of instances to each instance whose total score exceeds a predetermined threshold. We provide a computer device configured in such a way.

[0017] This computer device may be configured to perform each of the possible embodiments for processing proof images as described above.

[0018] Naturally, the various features, variations, and embodiments of the present invention may be combined with each other in various combinations, provided that they are not contradictory or mutually exclusive.

[0019] Other features and advantages of the present invention will become apparent from the following description with reference to the accompanying drawings illustrating examples of embodiments not intended to be limiting. [Brief explanation of the drawing]

[0020] [Figure 1] A schematic representation of preferred embodiments of a computer device and server that implement the method according to the present invention is shown. [Figure 2] Illustrate the steps of an embodiment of a method for processing a proof image showing a plurality of potential traces according to the present invention.

[0021] FIG. 1 schematically represents a preferred embodiment of a computer device 1 for processing a proof image and a server 2 for implementing the method according to the present invention.

[0022] The computer device 1 for processing a proof image includes a processor 10, a communication interface 12, and a memory 14. Further, the computer device 1 may include an image sensor 16 for displaying fingerprints.

[0023] The computer device 1 may be a device (e.g., a microcomputer, a workstation, or a lightweight portable device).

[0024] The processor 10 is configured to implement the steps of the proof image processing method described later. The processor 10 may have any structure. The processor includes one or more cores, and each core is configured to execute the instructions of a computer program to implement the above-described steps. It will be found later that this program uses a semantic segmentation neural network.

[0025] The communication interface 12 is suitable for the computer device 1 for processing a proof image to receive the proof image to be processed and communicate with the server 2. The communication interface 12 is any type of communication interface. For example, the communication interface 12 may be wired (Ethernet) or wireless using any communication protocol (Wi-Fi, Bluetooth, etc.).

[0026] Memory 14 is suitable for storing data operated or generated by the processor 10. Memory 14 may be any type of memory. Conventionally, memory 14 includes volatile memory for temporarily storing data and non-volatile memory for permanently storing data (i.e., in a way that retains data when the power to the non-volatile memory is turned off).

[0027] Memory 14 is particularly suitable for storing proof images that are received by computing device 1 and intended to be processed by processor 10.

[0028] Furthermore, the memory 14 stores computer program instructions, and some of these computer program instructions are designed to implement a method for processing proof images according to the present invention, as typically described with reference to Figure 2, when these instructions are executed by the processor 10.

[0029] Optionally, the image sensor 16 may include a transparent surface that functions as a support for the finger or palm, for example, to stabilize the finger or palm, and thus clearly show the fingerprint of the finger or the palm image of the palm in the image provided by the sensor.

[0030] The image sensor 16 can acquire a reference nipple trace.

[0031] Server 2 stores a database containing reference images that show reference nipple traces for previously registered reference individuals.

[0032] In this embodiment, computer device 1 is located away from server 2, as shown in Figure 1. However, in another embodiment, computer device 1 can constitute server 2.

[0033] In the explanation below, "fingerprint" refers to a papillary fingerprint, "latent fingerprint" refers to a latent papillary fingerprint, and "reference fingerprint" refers to a reference papillary fingerprint.

[0034] Here, the reference fingerprint is a fingerprint intended to be acquired in a controlled and autonomous manner using biometric acquisition means (e.g., image sensor 16).

[0035] Figure 2 illustrates the steps of an embodiment of a method for processing proof images showing multiple latent traces according to the present invention.

[0036] This method is carried out by computer device 1.

[0037] Here, the term "proof image" means a photograph of one or more papillary traces left in a culture medium by the addition of a region of skin having papillary protrusions, i.e., a photograph of one or more latent fingerprints.

[0038] According to step E102, which involves generating a semantic segmentation map, computer device 1 generates a semantic segmentation map from the test image using a neural network.

[0039] Typically, the test image is received by the computer device 1 via the communication interface 12 (a step not shown) and then stored in the computer device 1's memory 14.

[0040] Various semantic segmentation neural networks are known to those skilled in the art. As a non-existent example, semantic segmentation can be performed using convolutional neural networks, such as those described in Damien Fourure's doctoral dissertation from the University of Lyon, titled "Convolutional neural networks for semantic segmentation and colour invariant learning," which was publicly presented on December 12, 2017.

[0041] The semantic segmentation map includes, for each block that forms part of multiple blocks in the test image, a score associated with the block indicating the probability of presence in the test image of a latent fingerprint passing through that block.

[0042] Typically, for each block that forms part of a set of blocks in a test image, the score is a real number between 0 and 1 (i.e., 0 and 1) with a score associated with a block that tends to be 1 if the latent fingerprint passes through the block associated with the score, and 0 if the latent fingerprint does not pass through the block associated with the score.

[0043] In this text, an image “block” constitutes a contiguous region of the image that includes one or more pixels of the image. In a preferred embodiment, each conceivable block is actually a pixel of the test image. In this case, if multiple blocks of the test image are a set of blocks that include the entire test image, then multiple blocks of the test image are a set of pixels of the test image.

[0044] Next, the method includes a step (step E104) in which the computing device 1 generates a binary map from a semantic segmentation map.

[0045] The binary map estimates whether each block in a proof image represents part of a latent fingerprint.

[0046] The binary map includes a binary indicator associated with each block that forms part of a set of multiple blocks in the test image.

[0047] Typically, during this step (step E104) in which a binary map is generated, the computer device 1 binarizes the semantic segmentation map obtained in the step (step E102) in which a semantic segmentation map is generated, using a predetermined threshold.

[0048] For example, if each score is a real number between 0 and 1, as described above, the threshold may have a value of 0.5.

[0049] In this case, the binary indicator associated with the block is, -If the score associated with this block in the semantic segmentation map is below the threshold, 0 - If the score associated with this block in the semantic segmentation map exceeds the threshold, 1 It may have a value of . A binary indicator of 0 means that the associated block does not show a latent fingerprint portion, while a binary indicator of 1 means that the associated block shows a latent fingerprint portion.

[0050] Next, the method includes the step (step E106) of detecting multiple instances of a latent fingerprint from a binary map.

[0051] During this step (step E106), computer device 1 divides the potential fingerprint from the proof image into individual instances.

[0052] The step of detecting multiple instances of a latent fingerprint from a binary map consists of segmenting the latent fingerprint from the binary map by instance.

[0053] Conveniently, the step of detecting multiple latent fingerprint instances from a binary map (step E106) includes a substep (substep SE106_2) of applying watershed segmentation to the binary map by sliding window'').

[0054] Therefore, the method can better separate potential fingerprints that come into contact with each other.

[0055] A sliding window typically has an area of ​​u × u blocks (where u has a value determined as follows: u = 256 × r (where r is the image resolution of the block per inch)).

[0056] For example, a sliding window has a 256 x 256 pixel area for a proof image with a resolution of 500 pixels per inch (abbreviated as dpi), where in this example one block is one pixel.

[0057] In another example, the sliding window has a 512x512 pixel area relative to a proof image with a resolution of 1000 dpi.

[0058] Therefore, watershed segmentation focuses on detecting small latent traces, typically latent fingerprints. Furthermore, this window size speeds up proof image calculation while avoiding oversegmentation of latent fingerprints.

[0059] Various watershed segmentation algorithms are known to those skilled in the art. For example, a method may perform watershed segmentation as described in the paper "Watershed Segmentation Algorithm Based on Morphological Gradient Reconstruction," Baoan Han, 2015 2nd International Conference on Information Science and Control Engineering.

[0060] Watershed segmentation uses a binary map from a geographical perspective. This segmentation considers the binary map as topographic relief, where the distance from the nearest block that does not show part of the latent fingerprint to the block that shows part of the latent fingerprint represents the elevation of the block showing part of the latent fingerprint. Blocks that do not show part of the latent fingerprint have an elevation of zero. Next, watershed segmentation explores the watershed of this relief and separates different regions of the binary map. Each region obtained from the binary map corresponds to a watershed.

[0061] Similarly, the step of detecting multiple instances of latent fingerprints from a binary map (step E106) includes two other substeps. In this case, the step of detecting multiple instances of latent fingerprints from a binary map (step E106) includes a substep of obtaining another binary map (substep SE106_4), and then a substep of applying linked component analysis to the other binary map (substep SE106_6).

[0062] During the substep (substep SE106_4) to obtain another binary map, computer device 1 obtains another binary map from the binary map.

[0063] According to the first example, other binary maps are binary maps.

[0064] In the second example, another binary map is obtained by eroding a binary map with a predetermined region, and a block of the proof image showing a portion of the latent trace that has a region smaller than the predetermined region is removed.

[0065] A given area of ​​the erosion window is typically a u × u block (where u has a value determined as follows: u = 200 × r (where r is the resolution of the test image per inch of the block)).

[0066] A given area of ​​the erosion window is, for example, 200 x 200 pixels for a proof image with a resolution of 500 dpi, where in this example, a block is a pixel.

[0067] In another example, the erosion window has a 400x400 pixel area relative to a proof image with a resolution of 1000 dpi.

[0068] A region is considered smaller than another region, and this other region is defined by a window if it can contain a region within it, i.e., if it can contain a region within the window that defines the other region.

[0069] Due to the erosion of the binary map, analysis using connected components can be focused on detecting large latent traces, typically latent palm prints.

[0070] Next, during the substep (substep SE106_6) in which linked component analysis is applied to the other binary maps, computer device 1 applies linked component analysis to the other binary maps.

[0071] Related component analysis enables superior latent palmitate segmentation compared to watershed segmentation.

[0072] Various connected component analysis algorithms are known to those skilled in the art. For example, the method may perform connected component analysis as described in "The connected-component labelling problem: A review of state-of-the-art algorithms," Lifeng He et al., Elsevier: Pattern Recognition Volume 70 (2017), pages 25-43.

[0073] Linked component analysis connects all adjacent blocks to a given block that represents a portion of a latent fingerprint. Therefore, linked component analysis groups the connected (i.e., adjacent) blocks that represent portions of the latent fingerprint.

[0074] In the embodiment shown in Figure 2, the step of detecting multiple instances of latent fingerprints from a binary map (step E106) includes a substep of applying watershed segmentation (substep SE106_2), a substep of obtaining another binary map (substep SE106_4), and a substep of applying linked component analysis to the other binary map (substep SE106_6).

[0075] Depending on the method, one or more potential fingerprints and / or palm prints can be detected simultaneously and efficiently.

[0076] The substep (substep SE106_2) for applying watershed segmentation and watershed line segmentation is, for example, -By a processor including several cores, at least one core configured to perform a substep (substep SE106_2) to apply watershed segmentation, and at least one other core configured to perform a substep (substep SE106_4) to acquire another binary map and a substep (substep SE106_6) to apply connected component analysis to the other binary map, and / or -A number of processors, at least one processor configured to perform a substep (substep SE106_2) to apply watershed segmentation, and at least one other processor configured to perform a substep (substep SE106_4) to acquire another binary map and a substep (substep SE106_6) to apply connected component analysis to the other binary map, This may be performed in parallel with the substep of obtaining another binary map (substep SE106_4) and the substep of applying linked component analysis to the other binary map (substep SE106_6).

[0077] Alternatively, the computer device 1 may sequentially perform the following substeps: applying watershed segmentation (substep SE106_2), obtaining another binary map (substep SE106_4), and applying linked component analysis to the other binary map (substep SE106_6).

[0078] In the first alternative embodiment, the step of detecting multiple instances of latent fingerprints from a binary map (step E106) may include a substep of applying watershed segmentation (substep SE106_2), but may not include a substep of obtaining another binary map (substep SE106_4) or a substep of applying linked component analysis to the other binary map (substep SE106_6).

[0079] In a second embodiment, the step of detecting multiple instances of latent fingerprints from a binary map (step E106) may include a substep of obtaining another binary map (substep SE106_4) and a substep of applying linked component analysis to the other binary map (substep SE106_6), but may not include a substep of applying watershed segmentation (substep SE106_2).

[0080] The method involves reducing multiple instances, that is, continuing with a step (step E108) that reduces multiple instances detected during the detection step (step E106).

[0081] The step of reducing multiple instances (step E108) includes a substep (substep SE108_2) that defines the discovery box.

[0082] During this substep (substep SE108_2) in which the detection box is defined, computer device 1 defines a detection box for each instance of a plurality of instances, the detection box being the smallest rectangular box that completely encloses this instance.

[0083] Advantageously, the next step of reducing multiple instances (step E108) may include a substep (substep SE108_4) in which computing device 1 removes one instance from multiple instances, i.e., removes an instance from multiple instances detected during the detection step (step E106), if the detection box contains a detection box of another instance from these multiple instances.

[0084] Conveniently, the step of reducing multiple instances (step E108) may include a substep of merging (substep SE108_6) after a substep of removing at least one instance from multiple instances (substep SE108_4).

[0085] During this merging substep (substep SE108_6), if the intersection of the detection box of the first instance and the detection box of the second instance exceeds a predetermined ratio of the combined detection box of the first instance and the detection box of the second instance, the computing device 1 merges the first instance and the second instance from among the multiple instances.

[0086] Merging the first and second instances yields a single instance, while the reduced number of instances includes the single instance but does not include either the first or second instance. The given ratio is typically 0.6.

[0087] In the embodiment shown in Figure 2, the step of reducing multiple instances (step E108) includes a substep of defining a detection box (substep SE108_2), a substep of deleting a detection box (substep SE108_4), and a substep of merging (substep SE108_6).

[0088] In the first alternative embodiment, the step of reducing multiple instances (step E108) may include a substep of defining a detection box (substep SE108_2) and a substep of deleting a detection box (substep SE108_4), but may not include a substep of merging (substep SE108_6).

[0089] In a second embodiment, the step of reducing multiple instances (step E108) may include a substep of defining a detection box (substep SE108_2) and a substep of merging (substep SE108_6), but may not include a substep of deleting a detection box (substep SE108_4). This second other embodiment is advantageous when the step of detecting multiple latent fingerprint instances from a binary map (step E106) follows the first other embodiment described above for this detection step (step E106).

[0090] After the step of reducing multiple instances (step E108), the multiple instances include a number of instances that is less than or equal to the number of instances detected during the detection step (step E106).

[0091] The step of reducing multiple instances can limit oversegmentation and improve processing efficiency by limiting the computations required in later steps.

[0092] The process continues with the step of obtaining the total score (step E109), in which computer device 1 obtains the total score from the semantic segmentation map for each instance of the multiple instances, that is, for each instance of the multiple instances that have been reduced by the reduction step (step E108) in this case.

[0093] During this step (step E109) ​​of obtaining the total score, computer device 1 obtains the total score by summing the scores associated with each block of the test image belonging to each instance of the multiple instances.

[0094] Typically, the total score of an instance is the average of the scores associated with blocks among multiple blocks in a test image that represent a portion of the instance's potential fingerprint.

[0095] The process continues with the step of obtaining a limited number of instances (step E110), and computer device 1 obtains a limited number of instances by limiting the number of instances to each instance whose total score exceeds another predetermined threshold.

[0096] Therefore, the total score can better eliminate false positives of potential fingerprint instances.

[0097] If the score is a real number between 0 and 1, and the total score of an instance is the average of the scores associated with a block among a plurality of blocks of a test image that represent a portion of the latent fingerprint of that instance, then preferably the other predetermined threshold is between 0.6 and 0.75 (i.e., including 0.6 and 0.75), for example, has a value of 0.6.

[0098] The restricted instances include all and only the instances among the multiple instances whose total score exceeds another predetermined threshold, i.e., all instances among the multiple instances whose total score exceeds another predetermined threshold and which have been reduced by the reduction step (step E108).

[0099] The method enables rapid and efficient detection of multiple potential fingerprints (typically through instance-based segmentation).

[0100] This process uses a semantic segmentation neural network followed by post-processing, and does not require an instance segmentation neural network.

[0101] Furthermore, the process may continue with an identification step (step E112) in which the computing device 1 individually compares at least one latent fingerprint instance from a limited number of instances with a reference fingerprint and verifies the match between each latent fingerprint of the at least one latent fingerprint instance and this one reference fingerprint.

[0102] Those skilled in the art will see that steps in this process can be omitted as long as other steps have the elements necessary for the execution of the step (e.g., multiple instances).

[0103] For example, the step of reducing multiple instances (step E108) may be omitted.

[0104] In this case, during this step (step E109) ​​of obtaining the total score, computer device 1 obtains the total score from the semantic segmentation map for each instance of the multiple instances, that is, for each instance of the multiple instances detected during the detection step (step E106).

[0105] Furthermore, the restricted multiple instances obtained during the step of obtaining restricted multiple instances (step E110) include all and only the instances among the multiple instances whose total score exceeds other predetermined thresholds, i.e., all instances among the multiple instances detected by the detection step (step E106) whose total score exceeds other predetermined thresholds.

[0106] Those skilled in the art will also understand that the steps and / or substeps of a process can be executed on the condition that each step and / or substep has elements necessary for the execution of the step (e.g., a binary map).

[0107] According to the first example, if the step of detecting multiple instances of latent fingerprints from a binary map (step E106) includes a substep of applying watershed segmentation (substep SE106_2), a substep of obtaining another binary map (substep SE106_4), and a substep of applying linked component analysis to the other binary map (substep SE106_6), - Before the substep to obtain another binary map (substep SE106_4) and the substep to apply connected component analysis to the other binary map (substep SE106_6), the substep to apply watershed segmentation (substep SE106_2) may be performed. - After the substeps of obtaining another binary map (substep SE106_4) and applying connected component analysis to the other binary map (substep SE106_6), a substep of applying watershed segmentation (substep SE106_2) may be performed. -After the substep of obtaining another binary map (substep SE106_4), and before the substep of applying connected component analysis to the other binary map (substep SE106_6), the substep of applying watershed segmentation (substep SE106_2) may be performed.

[0108] According to the second example, the step of obtaining the total score (step E109) ​​may be performed before the step of decreasing multiple instances (step E108).

[0109] In this case, when the step of reducing multiple instances (step E108) includes a substep of merging (SE108_6), and computing device 1 merges the first instance and the second instance among the multiple instances, the computing device merges the total score of the first instance with the total score of the second instance.

[0110] Typically, the computing device may relate to an instance resulting from the merge of a first instance and a second instance, the total score of an instance resulting from the merge having the highest value among the total scores of the first instance and the second instance, or the average (and possibly weight) of the total scores of the first instance and the second instance. The weight may be a function of the number of blocks in the first instance representing the latent fingerprint portion and the number of blocks in the second instance representing the latent fingerprint portion (the number of blocks in the instance representing the latent fingerprint portion can be determined from the binary map associated with this instance).

Claims

1. A method for processing proof images showing multiple latent traces, comprising the following steps performed by a computer device (1): - Step (E102) of generating a semantic segmentation map from the test image by a neural network of semantic segmentation maps, which includes a score associated with the block indicating the probability of the presence of a latent fingerprint passing through the block in the test image for each block forming part of a plurality of blocks in the test image, - A step (E104) of generating a binary map from the semantic segmentation map, wherein the binary map estimates whether each of the plurality of blocks in the test image represents a part of a latent fingerprint (E104), - The step of detecting multiple instances of latent traces from the binary map (E106), - A step (E109) ​​for each instance of the plurality of instances, wherein the total score is obtained by summing the scores associated with the blocks among the plurality of blocks of the test image belonging to the instance, - Step (E110) of obtaining a restricted set of instances by restricting each instance whose total score exceeds a predetermined threshold. A method that includes this.

2. A method for processing a proof image according to claim 1, wherein the total score of an instance is the average of the scores associated with a block among a plurality of blocks of the proof image that represent a portion of the latent fingerprint of the detected instance.

3. A method for processing a proof image according to claim 1 or 2, wherein the step of detecting multiple instances of a latent fingerprint (E106) includes a substep (SE106_2) of applying watershed segmentation to the binary map by a sliding window.

4. A method for processing a proof image according to any one of claims 1 to 3, wherein the step of detecting multiple instances of latent traces (E106) includes a substep (SE106_6) of applying component analysis-related component analysis to another binary map obtained from the binary map.

5. The step (E106) of detecting multiple instances of latent traces is, - A substep (SE106_4) to obtain the other binary map by eroding the binary map with an erosion window having a predetermined area, and to remove the block of the proof image that shows a latent fingerprint portion having an area smaller than the predetermined area. A method for processing a proof image according to any one of claims 1 to 4, further comprising:

6. Between the step (E106) of detecting multiple instances of latent traces and the step (E109) ​​of obtaining a total score for each instance of the multiple instances, the following substeps are taken: - A substep (SE108_2) that defines a detection box for each instance of the plurality of instances, wherein the detection box is a minimum rectangular box that completely contains the instance, - If the detection box is included in the detection box of another instance from the multiple instances, a substep (SE108_4) to remove the instance from the multiple instances, - If the intersection of the detection box of the first instance and the detection box of the second instance exceeds a predetermined ratio of the combined detection box of the first instance and the detection box of the second instance, a substep (SE108_6) is performed to merge the first instance and the second instance from the plurality of instances. A method for processing a proof image according to any one of claims 1 to 5, comprising the step (E108) of reducing the plurality of instances, including the above.

7. A method for processing a proof image according to any one of claims 1 to 6, further comprising an identification step (E112) of individually comparing at least one latent fingerprint instance of the restricted plurality of instances with a reference fingerprint and verifying a match between each latent fingerprint of the at least one latent fingerprint instance and the reference fingerprint.

8. A computer program that includes a code instruction that performs the method according to any one of claims 1 to 7 when the code instruction is executed by a computer.

9. A computer-readable storage medium for storing instructions that can be executed by a computer performing the method according to any one of claims 1 to 7.

10. A computer device (1) that processes proof images showing multiple latent traces, - A semantic segmentation map is generated from the evidence image by a neural network, and the semantic segmentation map includes, for each block forming part of a plurality of blocks of the evidence image, a score associated with the block indicating the probability of the presence of a latent fingerprint passing through the block in the evidence image. - A binary map is generated from the semantic segmentation map, and the binary map estimates whether each of the multiple blocks in the test image represents part of a latent fingerprint. - Multiple instances of latent fingerprints are detected from the aforementioned binary map, - For each instance of the plurality of instances, obtain a total score from the semantic segmentation map, and obtain the total score by summing the scores associated with the blocks among the plurality of blocks of the test image belonging to the instance, - Obtain a limited set of instances by restricting each instance whose total score exceeds a predetermined threshold. A computer device (1) configured as follows.

Images