Model training method, vein feature extraction method, equipment, medium and product
By introducing anchor mechanism and generalization loss classification method into convolutional neural networks, the problem of difficulty in extracting venous features in the existing technology is solved, and the recognition accuracy and adaptability are improved.
Patent Information
- Application Number
- CN202411915632.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-24
- Publication Date
- 2025-04-01
AI Technical Summary
Existing fingerprint venous images contain a lot of noise and irregular shadows, making it difficult to extract venous features and affecting the recognition accuracy.
By obtaining training data from the data set, input training venous images into the convolutional neural network, using the anchor mechanism to generate predicted venous features, and compute generalization loss for classification, finally determining the loss function to update the network parameters, improving the ability to extract venous features.
The model's ability to extract venous features was improved, the classification evaluation criteria were optimized, the regression loss was reduced, the ability to adapt to venous features was enhanced, and missed and missed detection was reduced.
Smart Images
Figure CN119741740A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of biometric identification technology, and in particular to a model training method, a vein feature extraction method, equipment, medium and product. Background Art
[0002] With the digital transformation of financial services, data security and privacy protection are increasingly valued by users. As a novel biometric identification method, the vein image-based identity recognition method can well adapt to market demand under the above background and has become a new hot spot in biometric identification research in recent years.
[0003] Taking fingerprint vein images as an example, the vein image-based identity recognition method is based on the scientific principle that the blood flowing in human fingers can absorb light of a specific wavelength. It uses the fingerprint vein image obtained after near-infrared rays penetrate the finger to perform personal identification. Compared with traditional fingerprint or face recognition methods, it has the advantages of being difficult to forge, high recognition accuracy, and non-contact. However, affected by various factors (such as the collection environment), the existing fingerprint vein images contain a lot of noise and irregular shadows, which makes it difficult to effectively extract the vein features with distinguishing information, thereby affecting the recognition accuracy. Summary of the invention
[0004] The present invention provides a model training method, a vein feature extraction method, equipment, medium and product, which can improve the model's ability to extract vein features and provide a basis for subsequent identification of user identities.
[0005] According to one aspect of the present invention, a model training method is provided, comprising: obtaining a set of training data from a data set, wherein the training data comprises a training vein image and its corresponding calibration feature information, and the calibration feature information is used to characterize the actual vein features in the training vein image; inputting the training vein image into a convolutional neural network to obtain training anchor point information of the training vein image, wherein the training anchor point information is used to characterize the predicted vein features in the training vein image; respectively calculating the generalization loss between each training anchor point information and the calibration feature information, and classifying the training anchor point information according to the generalization loss; determining a loss function according to the classification result, the training anchor point information and the calibration feature information; if the training end condition is not met, updating the parameters of the convolutional neural network according to the loss function, and re-obtaining a set of training data from the data set, and returning to execute the step of inputting the training vein image into the convolutional neural network to obtain the training anchor point information of the training vein image; if the training end condition is met, using the current convolutional neural network as a target model, wherein the target model is used to extract the vein features of the user.
[0006] Optionally, the training vein image is input into a convolutional neural network to obtain training anchor point information of the training vein image, including: extracting features of the training vein image to obtain an original feature map; the convolution layer of the convolutional neural network determines a number of anchor points based on the original feature map, and predicts the position information and shape information of each anchor point; the pooling layer of the convolutional neural network maps the position information and shape information of each anchor point to the original feature map to determine the training anchor point information of the training vein image.
[0007] Optionally, for any training anchor point information, the generalization loss between the training anchor point information and the calibration feature information is calculated, including: obtaining the anchor box corresponding to the training anchor point information, the calibration box corresponding to the calibration feature information, and the minimum bounding box of the anchor box and the calibration box; according to the intersection area of the anchor box and the calibration box, the union area of the anchor box and the calibration box, and the area of the minimum bounding box, the generalization loss between the training anchor point information and the calibration feature information is calculated.
[0008] Optionally, the generalization loss between the training anchor point information and the calibration feature information is equal to the difference between the first ratio and the second ratio; wherein the first ratio is equal to the ratio of the intersection area of the anchor box and the calibration box to the union area of the anchor box and the calibration box; the second ratio is equal to the ratio of the difference between the area of the minimum bounding box and the union area of the anchor box and the calibration box to the area of the minimum bounding box.
[0009] Optionally, when the generalization loss is greater than or equal to a preset threshold, the classification result of the training anchor point information is a positive sample; when the generalization loss is less than the preset threshold, the classification result of the training anchor point information is a non-positive sample.
[0010] Optionally, a loss function is determined based on the classification results, training anchor point information and calibration feature information, including: determining a classification loss sub-function and a regression loss sub-function respectively based on the classification results, training anchor point information and calibration feature information; and determining a loss function based on weight parameters, the classification loss sub-function and the regression loss sub-function.
[0011] According to another aspect of the present invention, a vein feature extraction method is provided, comprising: acquiring a target vein image of a user; inputting the target vein image into a target model to obtain target anchor point information of the target vein image, wherein the target model is obtained using the model training method of any embodiment of the present invention; and extracting the vein features of the user based on the target anchor point information.
[0012] Optionally, after extracting the user's vein features, the method further includes: storing the user's vein features as the user's standard features; or comparing the user's vein features with pre-stored user's standard features to obtain the user's identity recognition results.
[0013] According to another aspect of the present invention, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the model training method of any embodiment of the present invention, or can execute the vein feature extraction method of any embodiment of the present invention.
[0014] According to another aspect of the present invention, a computer-readable storage medium is provided, which stores computer instructions, and the computer instructions are used to enable a processor to implement the model training method of any embodiment of the present invention, or implement the vein feature extraction method of any embodiment of the present invention when executed.
[0015] According to another aspect of the present invention, a computer program product is provided. The computer program product includes a computer program. When the computer program is executed by a processor, it implements the model training method of any embodiment of the present invention, or implements the vein feature extraction method of any embodiment of the present invention.
[0016] The technical solution of the embodiment of the present invention obtains a set of training data from a data set, and inputs the training vein image included in the training data into a convolutional neural network to obtain the training anchor information of the training vein image; then the generalization loss between each training anchor information and the calibration feature information corresponding to the training vein image is calculated to achieve the classification of the training anchor information; further, the loss function is determined based on the classification result, the training anchor information and the calibration feature information, so as to update the parameters of the convolutional neural network based on the loss function, continuously train the convolutional neural network, and finally use the trained convolutional neural network as the target model to extract the user's vein features. On the one hand, the model training method introduces an anchor mechanism in the convolutional neural network, which can enhance the learning ability of the convolutional neural network, thereby improving the target model's ability to extract vein features. On the other hand, the training anchor information is classified by the generalization loss between the training anchor information and the calibration feature information, and then the loss function is constructed for the parameter update of the convolutional neural network. In this way, the classification evaluation criteria can be optimized to better reflect the degree of fit between the training anchor information and the calibration feature information, thereby reducing the regression loss of the convolutional neural network and enhancing the adaptability of the convolutional neural network to the vein features, thereby reducing missed detection and false detection. In addition, the trained target model can accurately extract the user's vein features, providing a basis for subsequent identification of user identities, so as to promote the development of financial security construction.
[0017] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0019] Figure 1 It is a flowchart of a model training method provided by an embodiment of the present invention;
[0020] Figure 2 is a flow chart of another model training method provided by an embodiment of the present invention;
[0021] Figure 3 is a schematic diagram of an anchor point frame, a calibration frame, and a minimum bounding box of the anchor point frame and the calibration frame provided by an embodiment of the present invention;
[0022] Figure 4 is a flow chart of a vein feature extraction method provided by an embodiment of the present invention;
[0023] Figure 5 is a structural schematic diagram of a model training device provided by an embodiment of the present invention;
[0024] Figure 6 is a structural schematic diagram of a vein feature extraction device provided by an embodiment of the present invention;
[0025] Figure 7 is a schematic structural diagram of another vein feature extraction device provided by an embodiment of the present invention;
[0026] Figure 8 It is a structural schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0027] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.
[0028] It should be noted that the terms "first", "second", "training", "target", "original", "calibrated", "actual", "predicted", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0029] It should also be noted that the relevant information (including but not limited to user device information, user personal information, user vein images, etc.) and data (including but not limited to data for display, data for analysis, etc.) involved in the present invention are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data comply with relevant laws, regulations and standards in relevant regions.
[0030] Figure 1 It is a flow chart of a model training method provided by an embodiment of the present invention. This embodiment can be applied to the scenario of training a target model that can extract the user's vein features. The method can be executed by a model training device, which can be implemented in the form of hardware and / or software. The model training device can be configured in an electronic device (such as a computer device, a server, etc.).
[0031] like Figure 1 As shown, the method includes:
[0032] S110, obtaining a set of training data from the data set, wherein the training data includes a training vein image and its corresponding calibration feature information, and the calibration feature information is used to characterize actual vein features in the training vein image.
[0033] In the present invention, a data set is a training data set used for model training. A set of training data in a data set is the data required by a convolutional neural network during one training. The data set includes a number of training data in total. A set of training data is a part of a number of training data, and the number of training data can be one or more.
[0034] A training data usually includes a training vein image and calibration feature information corresponding to the training vein image, and the calibration feature information is used to characterize the actual vein features in the training vein image. The vein image can be a fingerprint vein image or a palm vein image.
[0035] In one embodiment, different training vein images may correspond to different users, that is, each training vein image is a vein image collected from a different user. In this way, the storage capacity occupied by the data set can be reduced, and the privacy of the user can be protected.
[0036] The calibration feature information may be the coordinate information of the actual vein feature in the training vein image, or the range information of the actual vein feature in the training vein image. When a training vein image corresponds to multiple calibration feature information, the multiple calibration feature information may be arranged in a certain order (such as from left to right, from top to bottom, from small to large, etc.) to facilitate subsequent data processing.
[0037] S120, inputting the training vein image into a convolutional neural network to obtain training anchor point information of the training vein image, wherein the training anchor point information is used to characterize predicted vein features in the training vein image.
[0038] Convolutional Neural Networks (CNN) is a deep learning model that can automatically learn feature representation from data by simulating the visual perception mechanism of the human brain. In the present invention, by introducing an anchor mechanism in the convolutional neural network, a possible object bounding box (i.e., the anchor box corresponding to the anchor point) is generated in the prediction process of the convolutional neural network, and then the position and category of the object bounding box are predicted to achieve the purpose of extracting vein features from the training vein image.
[0039] The training anchor point information may be the coordinate information of the predicted vein feature in the training vein image, or may be the range information of the predicted vein feature in the training vein image.
[0040] In a possible implementation, the number of training anchor point information of a training vein image is the same as the number of calibration feature information corresponding to the training vein image, that is, one training anchor point information corresponds to one calibration feature information.
[0041] In another possible implementation, the number of training anchor point information of a training vein image is greater than the number of calibration feature information corresponding to the training vein image, that is, one calibration feature information corresponds to multiple training anchor point information. Compared with the previous implementation, the number of training anchor point information increases, which can improve the feature analysis accuracy of each anchor point.
[0042] S130, respectively calculating the generalization loss between each training anchor point information and the calibration feature information, and classifying the training anchor point information according to the generalization loss.
[0043] After obtaining the training anchor point information of the training vein image, it is also necessary to match the training anchor point information with the calibration feature information to achieve classification of the training anchor point information.
[0044] Specifically, matching the training anchor point information with the calibration feature information can be achieved by using a matching algorithm (such as the Intersection over Union (IoU) algorithm), or by calculating the generalization loss between the training anchor point information and the calibration feature information. Compared with the overlap algorithm, calculating the generalization loss between the training anchor point information and the calibration feature information can avoid the problem of being unable to perform error feedback that may be caused by the existing matching algorithm, thereby better reflecting the degree of fit between the training anchor point information and the calibration feature information. Therefore, the present invention preferably uses generalization loss as a classification evaluation criterion.
[0045] Generalization Loss refers to the performance of the model on unseen data, also known as the generalization ability of the model. It is a key indicator to measure the performance of the model in practical applications, reflecting the prediction ability of the model on new data. The present invention measures the generalization loss by training the anchor box corresponding to the anchor information, the calibration box corresponding to the calibration feature information, and the minimum bounding box of the anchor box and the calibration box.
[0046] For example, the generalization loss between the training anchor information and the calibration feature information can be calculated based on the intersection area of the anchor box and the calibration box, the union area of the anchor box and the calibration box, and the area of the minimum bounding box.
[0047] Optionally, after the training anchor point information is classified according to the generalization loss, corresponding classification labels may be assigned to the training anchor point information so as to quickly determine the classification results during subsequent training.
[0048] S140: Determine a loss function according to the classification result, the training anchor point information and the calibration feature information.
[0049] Specifically, the predicted feature vector can be determined according to the classification result and the training anchor point information, and the calibrated feature vector can be determined according to the calibration feature information; and then the loss function can be determined based on the predicted feature vector and the calibrated feature vector.
[0050] By converting information into vectors, linear transformation of data can be achieved, thereby realizing end-to-end training, accelerating model training, and improving training results.
[0051] S150: Determine whether the training end condition is met. If yes, execute S170; if no, execute S160.
[0052] In one embodiment, the training end condition includes at least one of the following three conditions:
[0053] Condition 1: The value of the loss function is less than or equal to the preset loss;
[0054] Condition 2: The loss function converges;
[0055] Condition 3: The current number of training times is equal to the maximum number of training times.
[0056] When the training end condition includes at least two of the above conditions, as long as any one of the conditions is met, it is confirmed that the model training meets the training end condition.
[0057] S160. Update the parameters of the convolutional neural network according to the loss function, reacquire a set of training data from the data set, and return to execute S120.
[0058] When the training end condition is not met, it means that the current convolutional neural network has not been trained. At this time, the parameters of the convolutional neural network are updated according to the loss function, and a set of training data is reacquired from the data set, and the process returns to execute S120 until the training end condition is met.
[0059] S170. Using the current convolutional neural network as a target model, wherein the target model is used to extract vein features of the user.
[0060] When the training end condition is met, it means that the current convolutional neural network has been trained. At this time, the current convolutional neural network can be used as the target model, which is used to extract the user's vein features.
[0061] Optionally, after taking the current convolutional neural network as the target model, the target model can also be verified using a verification data set so that the parameters of the target model can be further fine-tuned when the verification result is not met to improve the vein feature extraction effect of the target model.
[0062] The technical solution of the embodiment of the present invention obtains a set of training data from a data set, and inputs the training vein image included in the training data into a convolutional neural network to obtain the training anchor information of the training vein image; then the generalization loss between each training anchor information and the calibration feature information corresponding to the training vein image is calculated to achieve the classification of the training anchor information; further, the loss function is determined based on the classification result, the training anchor information and the calibration feature information, so as to update the parameters of the convolutional neural network based on the loss function, continuously train the convolutional neural network, and finally use the trained convolutional neural network as the target model to extract the user's vein features. On the one hand, the model training method introduces an anchor mechanism in the convolutional neural network, which can enhance the learning ability of the convolutional neural network, thereby improving the target model's ability to extract vein features. On the other hand, the training anchor information is classified by the generalization loss between the training anchor information and the calibration feature information, and then the loss function is constructed for the parameter update of the convolutional neural network. In this way, the classification evaluation criteria can be optimized to better reflect the degree of fit between the training anchor information and the calibration feature information, thereby reducing the regression loss of the convolutional neural network and enhancing the adaptability of the convolutional neural network to the vein features, thereby reducing missed detection and false detection. In addition, the trained target model can accurately extract the user's vein features, providing a basis for subsequent identification of user identities, so as to promote the development of financial security construction.
[0063] Figure 2 FIG. 1 is a flow chart of another model training method provided by an embodiment of the present invention. Based on the above embodiment, this embodiment provides a specific method for determining training anchor information, a specific method for calculating generalization loss, and a specific method for determining loss function. Figure 2 As shown, the method includes:
[0064] S201. Acquire a set of training data from a data set, wherein the training data includes a training vein image and corresponding calibration feature information, and the calibration feature information is used to characterize actual vein features in the training vein image.
[0065] The data set is a training data set used for model training. Specifically, a set of training data can be obtained randomly from the data set, or a set of training data can be obtained from the data set according to a preset acquisition rule. A set of training data is the data required by the convolutional neural network during one training. The data set includes a number of training data. A set of training data is a part of a number of training data, and the number can be one or more.
[0066] A training data usually includes a training vein image and calibration feature information corresponding to the training vein image, and the calibration feature information is used to characterize the actual vein features in the training vein image. The vein image can be a fingerprint vein image or a palm vein image.
[0067] Exemplarily, the calibration feature information may be coordinate information of a calibration frame of an actual vein feature in a training vein image.
[0068] It should be noted that both the calibration feature information and the training anchor point information can be represented in the form of four-dimensional group information, namely (x, y, w, h). Among them, (x, y) represents the center position coordinates of the calibration frame corresponding to the calibration feature information / anchor frame corresponding to the training anchor point information, w represents the width of the calibration frame / anchor frame, and h represents the height of the calibration frame / anchor frame.
[0069] S202: Extract features from the training vein image to obtain an original feature map.
[0070] The network structure of the convolutional neural network may include an input layer, a convolutional layer, a pooling layer, a loss layer, etc. After the training vein image is input into the convolutional neural network, the input layer may first perform feature extraction on the training vein image to obtain an original feature map.
[0071] S203, the convolution layer of the convolutional neural network determines a number of anchor points based on the original feature map, and predicts the position information and shape information of each anchor point.
[0072] Since the training anchor point information is represented in the form of four-dimensional group information (x, y, w, h), according to the joint distribution formula p(x, y, w, h|I) = p(x, y|I)p(w, h|x, y, I), the position and shape of the anchor point in the training vein image I follow the distribution conditional on I. From this, two rules can be inferred: one is the specific area of the target object in the training vein image I; the other is that the size and proportion of the target object are closely related to its location. Therefore, the convolutional layer of the convolutional neural network is designed to predict the position information and shape information of each anchor point.
[0073] Specifically, the prediction may be performed in the order of position prediction first and shape prediction second.
[0074] For position prediction, the purpose is to determine whether the anchor point is the center point of the region of interest (i.e., the vein feature area), which is a binary classification problem. The specific process is: first, a 1×1 convolution operation is performed on the original feature map, and then the Sigmoid function is used to convert different distances into probability values for each element. According to the threshold, some areas are selected as candidate center points of the region of interest (i.e., anchor points), so that 90% of the areas can be filtered and the same recall rate is maintained. Among them, the convolution operation is an operation of performing an inner product (element multiplication and summation) on the image and the filter matrix.
[0075] For shape prediction, since the position of the region of interest is not changed, there is no problem of mismatch between the region of interest and the original feature map. The specific process is: predict (w, h) at each position on the given original feature map, but the value range of (w, h) may be too large, and it is difficult to predict directly. Therefore, the following transformation is made: w = σ·s·e dw ,h=σ·s·e dh , the outputs of dw and dh are mapped to (w,h) according to the above formula in the shape prediction stage. Among them, s represents the step size and σ represents the empirical scale factor. This nonlinear transformation compresses the output space from [0,1000] to [-1,1]. First, the 1×1 convolution layer generates a feature map of two channels (including the values of dw and dh), and then w and h are transformed under the action of each conversion layer, and finally the optimal length and width are predicted.
[0076] Optionally, the empirical scaling factor σ is taken as 8.
[0077] In addition, after predicting the position and shape information of each anchor point, in order to adapt to regions of interest of different sizes, the size of the region of interest is set proportional to the size of the corresponding receptive field. By using deformable convolution, the effective range and shape of the feature map can be closer to the purpose.
[0078] S204, the pooling layer of the convolutional neural network maps the position information and shape information of each anchor point to the original feature map to determine the training anchor point information of the training vein image.
[0079] The training anchor point information is used to characterize the predicted vein features in the training vein images.
[0080] The purpose of the pooling layer of the convolutional neural network is to use pooling to fix the size of regions of interest of different sizes. Specifically, the pooling layer of the convolutional neural network maps the position information and shape information of each anchor point to the original feature map, divides the mapping area into units according to the output dimension size, and performs maximum pooling on each unit to fix the feature size to determine the training anchor point information of the training vein image.
[0081] Optionally, the pooling layer of the convolutional neural network can also convert the training anchor information and calibration feature information into feature vectors through a fully connected layer to facilitate the subsequent loss function determination, realize end-to-end training, accelerate network training, and improve detection accuracy.
[0082] S205: Obtain an anchor point frame corresponding to the training anchor point information, a calibration frame corresponding to the calibration feature information, and a minimum bounding box of the anchor point frame and the calibration frame.
[0083] After obtaining the training anchor point information of the training vein image, it is also necessary to match the training anchor point information with the calibration feature information to achieve classification of the training anchor point information. The present invention selects generalization loss as the classification evaluation criterion.
[0084] Generalization loss refers to the performance of the model on unseen data, also known as the generalization ability of the model. It is a key indicator to measure the performance of the model in practical applications, reflecting the prediction ability of the model on new data. The present invention measures the generalization loss by training the anchor box corresponding to the anchor information, the calibration box corresponding to the calibration feature information, and the minimum bounding box of the anchor box and the calibration box.
[0085] Figure 3 Schematic diagram of an anchor box, a calibration box, and a minimum bounding box of the anchor box and the calibration box provided by an embodiment of the present invention. Figure 3 As shown, the anchor box is box A, the calibration box is box B, box A and box B intersect but do not overlap, and the minimum bounding box of the anchor box A and the calibration box B is box C.
[0086] S206: Calculate the generalization loss between the training anchor point information and the calibration feature information according to the intersection area of the anchor point box and the calibration box, the union area of the anchor point box and the calibration box, and the area of the minimum bounding box.
[0087] Specifically, the generalization loss between the training anchor information and the calibration feature information is equal to the difference between the first ratio and the second ratio. The first ratio is equal to the ratio of the intersection area of the anchor box and the calibration box to the union area of the anchor box and the calibration box; the second ratio is equal to the ratio of the difference between the area of the minimum bounding box and the union area of the anchor box and the calibration box to the area of the minimum bounding box.
[0088] When the anchor box and the calibration box completely overlap, the above method can calculate that the generalization loss between the training anchor information and the calibration feature information is equal to 1; when the anchor box and the calibration box do not overlap at all, the above method can calculate that the generalization loss between the training anchor information and the calibration feature information is equal to -1. Therefore, the optimization of the evaluation criteria for classification is achieved, the regression loss of the convolutional neural network is reduced, and the adaptability of the convolutional neural network to the vein features is enhanced, thereby reducing missed detections and false detections. Compared with the existing method of using the IoU algorithm as the evaluation criterion, it avoids the problem of being unable to return errors when loss = 0, is more in line with the distance measurement index, and can better reflect the degree of fit between the training anchor information and the calibration feature information.
[0089] S207. Classify the training anchor point information according to the generalization loss.
[0090] Specifically, when the generalization loss is greater than or equal to a preset threshold, the classification result of the training anchor point information is a positive sample; when the generalization loss is less than the preset threshold, the classification result of the training anchor point information is a non-positive sample.
[0091] In one embodiment, the value of the preset threshold can be set according to actual needs, such as 0.6, 0.7, 0.8, 0.9, etc.
[0092] In addition, for non-positive samples, the present invention can further divide the non-positive samples into neutral samples and negative samples.
[0093] The purpose of classifying the training anchor information is to avoid model bias caused by too few positive samples. According to the classification results, selecting appropriate samples for training can improve the robustness and performance of the model.
[0094] Optionally, after the training anchor point information is classified according to the generalization loss, corresponding classification labels may be assigned to the training anchor point information so as to quickly determine the classification results during subsequent training.
[0095] S208. Determine a classification loss sub-function and a regression loss sub-function respectively according to the classification result, the training anchor point information and the calibration feature information.
[0096] S209: Determine a loss function according to the weight parameter, the classification loss sub-function and the regression loss sub-function.
[0097] In combination with steps S208 and S209, when determining the loss function, the predicted feature vector can be determined according to the classification result and the training anchor point information, and the calibrated feature vector can be determined according to the calibration feature information; and then the loss function is determined based on the predicted feature vector and the calibrated feature vector. By converting information to vectors, linear transformation of data can be achieved, thereby achieving end-to-end training, accelerating model training, and improving training effects.
[0098] For example, the loss function
[0099] Among them, i represents the number of the anchor point, p i Indicates that the training anchor point information corresponding to anchor point i is the prediction confidence of the vein feature, Indicates the confidence of the calibration feature information. When the training anchor point information is a positive sample, Equal to 1; otherwise is equal to 0. i , is a vector containing four parameters, t i represents the x coordinate of the anchor to the bounding box, Represents the coordinate vector corresponding to the calibration feature information.
[0100] is a classification loss subfunction, which is essentially the logarithmic loss of two categories. It is defined as
[0101]
[0102] is the regression loss subfunction, defined as Among them, R represents the robust loss function Defined as
[0103] For bounding box regression, the following parameterization is performed: Among them, x, y, w, and h represent the horizontal and vertical coordinates of the anchor box corresponding to the training anchor information, as well as its border width and border height; x a Indicates the horizontal coordinate of the anchor point corresponding to the training anchor point information. * ,y * ,w * ,h * They respectively represent the horizontal and vertical coordinates of the calibration box corresponding to the calibration feature information and its border width and border height.
[0104] N cls and N reg They are used to standardize the two sub-functions respectively; λ is a weight parameter used to coordinate the balance between the classification loss sub-function and the regression loss sub-function. Optionally, the value of λ is 10, which can achieve the maximum balance between the classification loss sub-function and the regression loss sub-function.
[0105] S210: Determine whether the training end condition is met. If yes, execute S212; if no, execute S211.
[0106] In one embodiment, the training end condition includes at least one of the following three conditions:
[0107] Condition 1: The value of the loss function is less than or equal to the preset loss;
[0108] Condition 2: The loss function converges;
[0109] Condition 3: The current number of training times is equal to the maximum number of training times.
[0110] When the training end condition includes at least two of the above conditions, as long as any one of the conditions is met, it is confirmed that the model training meets the training end condition.
[0111] S211. Update the parameters of the convolutional neural network according to the loss function, reacquire a set of training data from the data set, and return to execute S202.
[0112] When the training end condition is not met, it means that the current convolutional neural network has not been trained. At this time, the parameters of the convolutional neural network are updated according to the loss function, and a set of training data is reacquired from the data set, and the process returns to execute S202 until the training end condition is met.
[0113] S212. Using the current convolutional neural network as a target model, wherein the target model is used to extract vein features of the user.
[0114] When the training end condition is met, it means that the current convolutional neural network has been trained. At this time, the current convolutional neural network can be used as the target model, which is used to extract the user's vein features.
[0115] Optionally, after taking the current convolutional neural network as the target model, the target model can also be verified using a verification data set so that the parameters of the target model can be further fine-tuned when the verification result is not met to improve the vein feature extraction effect of the target model.
[0116] The technical solution of the embodiment of the present invention obtains a set of training data from a data set, and inputs the training vein image included in the training data into a convolutional neural network to obtain the training anchor information of the training vein image; then the generalization loss between each training anchor information and the calibration feature information corresponding to the training vein image is calculated to achieve the classification of the training anchor information; further, the loss function is determined based on the classification result, the training anchor information and the calibration feature information, so as to update the parameters of the convolutional neural network based on the loss function, continuously train the convolutional neural network, and finally use the trained convolutional neural network as the target model to extract the user's vein features. On the one hand, the model training method introduces an anchor mechanism in the convolutional neural network, which can enhance the learning ability of the convolutional neural network, thereby improving the target model's ability to extract vein features. On the other hand, the training anchor information is classified by the generalization loss between the training anchor information and the calibration feature information, and then the loss function is constructed for the parameter update of the convolutional neural network. In this way, the classification evaluation criteria can be optimized to better reflect the degree of fit between the training anchor information and the calibration feature information, thereby reducing the regression loss of the convolutional neural network and enhancing the adaptability of the convolutional neural network to the vein features, thereby reducing missed detection and false detection. In addition, the trained target model can accurately extract the user's vein features, providing a basis for subsequent identification of user identities, so as to promote the development of financial security construction.
[0117] Figure 4 This is a flow chart of a vein feature extraction method provided by an embodiment of the present invention. This embodiment is applicable to the scenario of extracting the vein features of a user using the target model obtained by the above embodiment. The method can be executed by a vein feature extraction device, which can be implemented in the form of hardware and / or software, and can be configured in an electronic device (such as an identity recognition device, etc.). Figure 4 As shown, the method includes:
[0118] S410: Acquire a target vein image of the user.
[0119] The target vein image is usually collected based on the near-infrared light environment. Near infrared light (NIR) is an electromagnetic wave between visible light (VIS) and mid-infrared light (MIR). According to the definition of ASTM (American Society for Testing and Materials), it refers to electromagnetic waves with a wavelength in the range of 780-2526nm, which is usually divided into two regions: near infrared short wave (780-1100nm) and near infrared long wave (1100-2526nm). In the present invention, the wavelength of near infrared light can be in the range of 780-2526nm; preferably, the wavelength of near infrared light can be in the range of 780-1100nm, to ensure that the target vein image with better details is collected.
[0120] S420, inputting the target vein image into the target model to obtain target anchor point information of the target vein image, wherein the target model is obtained by using the model training method described in the above embodiment.
[0121] S430: Extracting the user's vein features according to the target anchor point information.
[0122] Specifically, the vein features of the user may be stored in the form of feature vectors, and the target model may directly perform linear transformation on the target anchor point information to obtain the vein features of the user.
[0123] Optionally, after extracting the user's vein features, the user's vein features may be stored as the user's standard features; or the user's vein features may be compared with pre-stored user's standard features to obtain the user's identity recognition result.
[0124] Storing the user's vein features as the user's standard features, that is, the process of generating the user's standard features, provides a basis for the application of biometric identity recognition; and comparing the user's vein features with the pre-stored user's standard features is the specific application of biometric identity recognition, which can improve the accuracy of identity recognition.
[0125] The vein feature extraction method provided in the embodiment of the present invention provides a basis for subsequent identification of the user's identity by accurately extracting the user's vein features, thereby promoting the development of financial security construction.
[0126] Figure 5 Schematic diagram of a model training device provided by an embodiment of the present invention. Figure 5 As shown, the device includes: a data acquisition module 501 and a training module 502.
[0127] The data acquisition module 501 is used to acquire a set of training data from the data set, wherein the training data includes a training vein image and its corresponding calibration feature information, and the calibration feature information is used to characterize the actual vein features in the training vein image;
[0128] The training module 502 is used to input the training vein image into the convolutional neural network to obtain the training anchor point information of the training vein image, wherein the training anchor point information is used to characterize the predicted vein features in the training vein image; respectively calculate the generalization loss between each training anchor point information and the calibration feature information, and classify the training anchor point information according to the generalization loss; determine the loss function according to the classification result, the training anchor point information and the calibration feature information; if the training end condition is not met, update the parameters of the convolutional neural network according to the loss function, and notify the data acquisition module 501 to reacquire a set of training data from the data set; if the training end condition is met, use the current convolutional neural network as the target model, wherein the target model is used to extract the user's vein features.
[0129] Optionally, the training module 502 is specifically used to extract features from the training vein image to obtain an original feature map; the convolution layer of the convolutional neural network determines a number of anchor points based on the original feature map, and predicts the position information and shape information of each anchor point; the pooling layer of the convolutional neural network maps the position information and shape information of each anchor point to the original feature map to determine the training anchor point information of the training vein image.
[0130] Optionally, the training module 502 is specifically used to obtain the anchor box corresponding to the training anchor point information, the calibration box corresponding to the calibration feature information, and the minimum bounding box of the anchor box and the calibration box; calculate the generalization loss between the training anchor point information and the calibration feature information according to the intersection area of the anchor box and the calibration box, the union area of the anchor box and the calibration box, and the area of the minimum bounding box.
[0131] Optionally, the generalization loss between the training anchor point information and the calibration feature information is equal to the difference between the first ratio and the second ratio; wherein the first ratio is equal to the ratio of the intersection area of the anchor box and the calibration box to the union area of the anchor box and the calibration box; the second ratio is equal to the ratio of the difference between the area of the minimum bounding box and the union area of the anchor box and the calibration box to the area of the minimum bounding box.
[0132] Optionally, when the generalization loss is greater than or equal to a preset threshold, the classification result of the training anchor point information is a positive sample; when the generalization loss is less than the preset threshold, the classification result of the training anchor point information is a non-positive sample.
[0133] Optionally, the training module 502 is specifically used to determine the classification loss sub-function and the regression loss sub-function according to the classification results, training anchor point information and calibration feature information; and determine the loss function according to the weight parameters, the classification loss sub-function and the regression loss sub-function.
[0134] The model training device provided in the embodiment of the present invention can execute the model training method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.
[0135] Figure 6 Schematic diagram of the structure of a vein feature extraction device provided by an embodiment of the present invention. Figure 6 As shown, the device includes: an image acquisition module 601 and a feature extraction module 602.
[0136] An image acquisition module 601 is used to acquire a target vein image of a user;
[0137] The feature extraction module 602 is used to input the target vein image into the target model to obtain the target anchor point information of the target vein image, wherein the target model is obtained by the model training method described in the above embodiment; and extract the user's vein features according to the target anchor point information.
[0138] Optional, reference Figure 6 , Figure 7 60 is a schematic diagram of another vein feature extraction device provided by an embodiment of the present invention. The device also includes: a storage module 603 and an identity recognition module 604.
[0139] The storage module 603 is used to store the vein characteristics of the user as the standard characteristics of the user;
[0140] The identity recognition module 604 is used to compare the vein characteristics of the user with the pre-stored standard characteristics of the user to obtain the identity recognition result of the user.
[0141] The vein feature extraction device provided in the embodiment of the present invention can execute the vein feature extraction method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.
[0142] Figure 8 : is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workbenches, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices (such as helmets, glasses, watches, etc.) and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present invention described and / or required herein.
[0143] like Figure 8As shown, the electronic device 10 includes at least one processor 11, and a memory connected to the at least one processor 11, such as a read-only memory (ROM) 12, a random access memory (RAM) 13, etc., wherein the memory stores a computer program that can be executed by at least one processor, and the processor 11 can perform various appropriate actions and processes according to the computer program stored in the read-only memory (ROM) 12 or the computer program loaded from the storage unit 18 to the random access memory (RAM) 13. In the RAM 13, various programs and data required for the operation of the electronic device 10 can also be stored. The processor 11, the ROM 12, and the RAM 13 are connected to each other through a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.
[0144] A number of components in the electronic device 10 are connected to the I / O interface 15, including: an input unit 16, such as a keyboard, a mouse, etc.; an output unit 17, such as various types of displays, speakers, etc.; a storage unit 18, such as a disk, an optical disk, etc.; and a communication unit 19, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 19 allows the electronic device 10 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.
[0145] The processor 11 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The processor 11 executes the various methods and processes described above, such as the model training method / vein feature extraction method.
[0146] In some embodiments, the model training method / vein feature extraction method may be implemented as a computer program, which is tangibly contained in a computer-readable storage medium, such as a storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 10 via the ROM 12 and / or the communication unit 19. When the computer program is loaded into the RAM 13 and executed by the processor 11, one or more steps of the model training method / vein feature extraction method described above may be performed. Alternatively, in other embodiments, the processor 11 may be configured to execute the model training method / vein feature extraction method in any other appropriate manner (e.g., by means of firmware).
[0147] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.
[0148] Computer programs for implementing the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that when the computer program is executed by the processor, the functions / operations specified in the flow chart and / or block diagram are implemented. The computer program may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.
[0149] In the context of the present invention, a computer-readable storage medium may be a tangible medium that may contain or store a computer program for use by or in combination with an instruction execution system, device or equipment. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or equipment, or any suitable combination of the foregoing. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.
[0150] To provide interaction with a user, the systems and techniques described herein may be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and a pointing device (e.g., a mouse or trackball) through which the user can provide input to the electronic device. Other types of devices may also be used to provide interaction with the user; for example, the feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form (including acoustic input, voice input, or tactile input).
[0151] The systems and techniques described herein may be implemented in a computing system that includes backend components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes frontend components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such backend components, middleware components, or frontend components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), a blockchain network, and the Internet.
[0152] A computing system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The client and server relationship is generated by computer programs running on the corresponding computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services.
[0153] An embodiment of the present invention further provides a computer program product, including a computer program, which, when executed by a processor, implements the model training method and / or vein feature extraction method provided in any embodiment of the present invention.
[0154] In the process of implementation, the computer program product can be written in one or more programming languages or a combination thereof to perform the computer program code of the present invention, including object-oriented programming languages, such as Java, Smalltalk, C++, and conventional procedural programming languages, such as "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (for example, using an Internet service provider to connect through the Internet).
[0155] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the present invention can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution of the present invention can be achieved, and this document does not limit this.
[0156] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A model training method, characterized in that: include: Acquire a set of training data from a data set, wherein the training data includes a training vein image and corresponding calibration feature information, and the calibration feature information is used to characterize actual vein features in the training vein image; Inputting the training vein image into a convolutional neural network to obtain training anchor point information of the training vein image, wherein the training anchor point information is used to characterize predicted vein features in the training vein image; Respectively calculating the generalization loss between each of the training anchor point information and the calibration feature information, and classifying the training anchor point information according to the generalization loss; Determining a loss function according to the classification result, the training anchor point information and the calibration feature information; If the training end condition is not met, then according to the loss function, the parameters of the convolutional neural network are updated, and a set of training data is reacquired from the data set, and the step of inputting the training vein image into the convolutional neural network to obtain the training anchor point information of the training vein image is returned to be executed; If the training end condition is met, the current convolutional neural network is used as the target model, wherein the target model is used to extract the user's vein features.
2. The model training method according to claim 1, characterized in that: The step of inputting the training vein image into a convolutional neural network to obtain training anchor point information of the training vein image includes: Performing feature extraction on the training vein image to obtain an original feature map; The convolution layer of the convolutional neural network determines a number of anchor points based on the original feature map, and predicts the position information and shape information of each anchor point; The pooling layer of the convolutional neural network maps the position information and shape information of each anchor point to the original feature map to determine the training anchor point information of the training vein image.
3. The model training method according to claim 1, characterized in that: For any training anchor point information, calculating the generalization loss between the training anchor point information and the calibration feature information includes: Obtaining an anchor point frame corresponding to the training anchor point information, a calibration frame corresponding to the calibration feature information, and a minimum bounding box of the anchor point frame and the calibration frame; The generalization loss between the training anchor point information and the calibration feature information is calculated according to the intersection area of the anchor point box and the calibration box, the union area of the anchor point box and the calibration box, and the area of the minimum circumscribed box.
4. The model training method according to claim 3, characterized in that: The generalization loss between the training anchor point information and the calibration feature information is equal to the difference between the first ratio and the second ratio; Among them, the first ratio is equal to the ratio of the intersection area of the anchor box and the calibration box to the union area of the anchor box and the calibration box; the second ratio is equal to the ratio of the difference between the area of the minimum circumscribed frame and the union area of the anchor box and the calibration frame to the area of the minimum circumscribed frame.
5. The model training method according to claim 3, characterized in that: When the generalization loss is greater than or equal to a preset threshold, the classification result of the training anchor point information is a positive sample; when the generalization loss is less than the preset threshold, the classification result of the training anchor point information is a non-positive sample.
6. The model training method according to claim 1, characterized in that: The determining of the loss function according to the classification result, the training anchor point information and the calibration feature information includes: Determine a classification loss sub-function and a regression loss sub-function respectively according to the classification result, the training anchor point information and the calibration feature information; The loss function is determined according to the weight parameter, the classification loss sub-function and the regression loss sub-function.
7. A vein feature extraction method, characterized in that: include: Collecting the user's target vein image; Inputting the target vein image into a target model to obtain target anchor point information of the target vein image, wherein the target model is obtained by using the model training method according to any one of claims 1 to 6; The vein features of the user are extracted according to the target anchor point information.
8. The vein feature extraction method according to claim 7, characterized in that: After extracting the vein features of the user, the method further includes: storing the vein feature of the user as a standard feature of the user; or, The vein features of the user are compared with the pre-stored standard features of the user to obtain the identity recognition result of the user.
9. An electronic device, characterized in that: The electronic device comprises: at least one processor; and a memory in communication with the at least one processor; wherein, The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the model training method described in any one of claims 1 to 6, or can execute the vein feature extraction method described in claim 7 or 8.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, which are used to enable a processor to implement the model training method described in any one of claims 1 to 6 when executed, or to execute the vein feature extraction method described in claim 7 or 8.
11. A computer program product, characterized in that The computer program product comprises a computer program, which, when executed by a processor, implements the model training method according to any one of claims 1 to 6, or is capable of executing the vein feature extraction method according to claim 7 or 8.