Palm vein enhancement method based on bidirectional connection segmentation algorithm and model training method
By acquiring and processing infrared palm vein and RGB palm print images separately based on a bidirectional connectivity segmentation algorithm, and extracting and retaining subtle features, the problem of low palm vein recognition accuracy in existing technologies is solved, and higher recognition accuracy is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHENZHEN MAXVISION TECH
- Filing Date
- 2022-09-13
- Publication Date
- 2026-07-07
AI Technical Summary
In existing technologies, palm vein recognition methods based on CNN deep learning suffer from reduced accuracy during the recognition process. In particular, they cannot effectively preserve the features of small vein branches and capillaries, and are easily affected by palm print interference features.
Infrared palm vein images and RGB palm print images were acquired using a bidirectional connection segmentation algorithm. Palm vein and palm print information were extracted using the bidirectional connection segmentation algorithm, and palm print interference in palm vein information was removed using palm print information, while preserving the main vein features and features such as capillaries and branch veins.
It improves the accuracy of palm vein recognition, ensuring that the model focuses more on the effective information of palm veins during training and recognition, reduces the influence of palm print interference, and improves the accuracy of recognition.
Smart Images
Figure CN115527245B_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of biometric recognition technology, and more specifically, it relates to a palm vein enhancement method and model training method based on a bidirectional connection segmentation algorithm. Background Technology
[0002] In critical security areas such as customs, prisons, and banks, it is necessary to distinguish between internal and external personnel, as well as personnel with different access permissions. Currently, facial recognition is not secure enough, so more secure biometric features, such as palm veins, are often used for personnel identification. Common infrared palm vein acquisition devices include infrared fill lights and infrared cameras. The infrared fill lights illuminate the palm, and some of the infrared light is absorbed by the palm veins and capillaries, forming palm vein and capillary features in the infrared palm vein image. Some of the infrared light is reflected after shining on the skin and is then captured by the infrared camera. Thus, the infrared palm vein image acquired by the infrared camera contains palm vein and capillary features, as well as a small amount of palmprint interference features.
[0003] In existing technologies, a common method for palm vein feature extraction is based on CNN deep learning. However, when this method enhances palm vein features, it only obtains coarse vein features. Fine vein branches and capillary features are filtered out, resulting in the loss of effective information for palm vein recognition. Even so, some palm vein features with obvious patterns still interfere, leading to a decrease in the accuracy of palm vein recognition. Summary of the Invention
[0004] The purpose of this application is to provide a palm vein enhancement method and model training method based on a bidirectional connection segmentation algorithm, so as to solve the technical problem of reduced accuracy in the palm vein recognition process in the prior art.
[0005] To achieve the above objectives, the technical solution adopted in this application is: to provide a palm vein enhancement method based on a bidirectional connection segmentation algorithm, wherein the palm vein enhancement method based on the bidirectional connection segmentation algorithm includes:
[0006] Step S1: Acquire infrared palm vein images and RGB palm print images respectively;
[0007] Step S2: Extract palm vein information from infrared palm vein images using a bidirectional connection-based segmentation algorithm, and extract palm print information from RGB palm print images using a bidirectional connection-based segmentation algorithm.
[0008] Step S3: Remove palmprint interference information from palm vein information based on palmprint information.
[0009] Preferably, in step S2, the bidirectional connection segmentation algorithm includes the following steps:
[0010] Step S21: Extract features from the input image x through convolutional layers C1, C2, and C3 to obtain features C1(x), C2(x), and C3(x), respectively.
[0011] Step S22: Perform deconvolution operations on features C2(x) and C3(x) to obtain deconvolution results D1 and D2 respectively;
[0012] Step S23: Superimpose features C1(x), C2(x), and C3(x) with the deconvolution results D1 and D2 to obtain the level outputs zx1, zx2, and zx3 corresponding to different levels.
[0013] Preferably, in step S23, the calculation formulas for hierarchical output zx1, hierarchical output zx2, and hierarchical output zx3 are as follows:
[0014] zx1=C1(x)+D1,
[0015] zx² = C²(x) + D²,
[0016] zx3=C3(x).
[0017] Preferably, in step S22, the method for deconvolution operation includes the following steps:
[0018] Step S231, define the input image X for the deconvolution operation as:
[0019]
[0020] The convolution kernel K is:
[0021]
[0022] Step S232, output image X 、 The size is 5×5. The input image X is padded with 0s between every two elements, i.e.:
[0023]
[0024] Step S233, map the convolution kernel K to X 、 Performing a convolution operation yields the result Y of the deconvolution operation, i.e.:
[0025]
[0026] Preferably, after step S23, the bidirectional connection segmentation algorithm further includes the following steps:
[0027] Step S241: Perform deconvolution operations on the hierarchical output zx1 and hierarchical output zx2 to obtain deconvolution operation results D3 and D4 respectively;
[0028] Step S242: Superimpose the hierarchical outputs zx1, zx2, and zx3 with the deconvolution results D3 and D4 to obtain the hierarchical outputs zy1, zy2, and zy3 corresponding to different levels.
[0029] Preferably, in step S242, the calculation formulas for hierarchical output zx1, hierarchical output zx2, and hierarchical output zx3 are as follows:
[0030] zy1 = zx1,
[0031] zy2=zx2+D3,
[0032] zy3=zx3+D4.
[0033] Preferably, after step S242, the bidirectional connection segmentation algorithm further includes the following steps:
[0034] Step S25: Scaling operation is performed on the hierarchical outputs zy1, zy2, and zy3 to obtain a standard image with consistent resolution.
[0035] Preferably, in step S3, the method for removing palm print interference information from palm vein information based on palm print information includes:
[0036] Step S31: Obtain the intersection information of palm vein information and palm print information.
[0037] Step S32: Remove the overlapping information from the palm vein information.
[0038] This application also provides a model training method, which includes the following steps:
[0039] The palm vein enhancement method based on the bidirectional connection segmentation algorithm described above was used to obtain a large number of training datasets;
[0040] The training dataset is input into the palm vein recognition CNN model for training.
[0041] Preferably, the palm vein recognition CNN model is a palm vein recognition CNN model based on ResNet50.
[0042] The palm vein enhancement method based on bidirectional connectivity segmentation algorithm provided in this application, compared with the prior art, utilizes the different characteristics of image acquisition devices to acquire infrared palm vein images and RGB palmprint images respectively. First, the palm vein information and palmprint information are enhanced based on the bidirectional connectivity segmentation algorithm, preserving subtle features at different levels. Then, the information obtained after segmentation and enhancement is used to verify each other. By removing the corresponding palmprint interference information in the palm vein information based on the palmprint information, the main vein features, capillary and branch vein features are preserved. In this way, the model can focus more on the effective information of palm veins during training and recognition, thereby improving the palm vein recognition accuracy.
[0043] Compared with existing technologies, the model training method provided in this application adopts a palm vein enhancement method based on bidirectional connection segmentation algorithm to obtain a large training dataset. This training dataset retains subtle features at different levels. By removing the corresponding palm print interference information from the palm vein information based on the palm print information, the main vein features, capillary and branch vein features are retained. This makes the model more focused on the effective information of the palm vein during training and recognition, and improves the accuracy of the palm vein recognition CNN model in palm vein recognition. Attached Figure Description
[0044] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0045] Figure 1 A flowchart illustrating the palm vein enhancement method based on a bidirectional connectivity segmentation algorithm provided in this application embodiment;
[0046] Figure 2 Based on Figure 1 A schematic diagram illustrating the effect of infrared palm vein images acquired using the method described above;
[0047] Figure 3 Based on Figure 1 The diagram illustrates the effect of the method in the paper, which extracts palm vein information from infrared palm vein images using a bidirectional connection segmentation algorithm.
[0048] Figure 4 Based on Figure 1 The flowchart of the method in the paper, which extracts palm vein information from infrared palm vein images using a bidirectional connection segmentation algorithm, is shown in the figure.
[0049] Figure 5 Based on Figure 4 Another schematic diagram of the bidirectional connection segmentation algorithm in the diagram;
[0050] Figure 6 This is a flowchart illustrating the model training method provided in an embodiment of this application. Detailed Implementation
[0051] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.
[0052] It should be noted that when a component is referred to as being "fixed to" or "set on" another component, it can be directly on or indirectly on that other component. When a component is referred to as being "connected to" another component, it can be directly connected to or indirectly connected to that other component.
[0053] It should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.
[0054] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.
[0055] Please refer to the following: Figures 1 to 3 The palm vein enhancement method based on a bidirectional connectivity segmentation algorithm provided in this application will now be described. The palm vein enhancement method based on the bidirectional connectivity segmentation algorithm includes:
[0056] Step S1: Acquire infrared palm vein images and RGB palm print images respectively;
[0057] Step S2: Extract palm vein information from infrared palm vein images using a bidirectional connection-based segmentation algorithm, and extract palm print information from RGB palm print images using a bidirectional connection-based segmentation algorithm.
[0058] Step S3: Remove palmprint interference information from palm vein information based on palmprint information.
[0059] It is worth noting that in step S1, the method for acquiring infrared palm vein images can use an infrared palm vein acquisition device to acquire palm vein image information. For example, the infrared palm vein acquisition device includes an infrared fill light and an infrared camera. The infrared fill light provides infrared illumination to the palm, and part of the infrared light is absorbed by the palm veins and capillaries, forming palm vein features and capillary features in the infrared palm vein image, that is, darker lines in the infrared palm vein image. Part of the infrared light is reflected to the infrared camera after illuminating the skin. Thus, the infrared palm vein image acquired by the infrared camera contains palm vein features, capillary features, and a small amount of palm print interference features.
[0060] In step S1, the method for acquiring RGB palmprint images can use a conventional RGB camera to simultaneously acquire RGB palmprint images based on the same palm. Since RGB cameras can only capture visible light, while infrared light is invisible light, the RGB palmprint images do not contain palm vein features or capillary features, but have clear palmprint information.
[0061] In step S2, palm vein information from the infrared palm vein image is extracted using a bidirectional connection-based segmentation algorithm, and palmprint information from the RGB palmprint image is extracted using the same algorithm. Simultaneously, the palm vein information from both the infrared and RGB palmprint images is enhanced, thus enabling the acquisition of feature information from different levels. This is because different levels correspond to different features; for example, in the infrared palm vein image, vein features are coarse while capillary features are fine, belonging to different levels.
[0062] Existing segmentation models, such as the PSPNet semantic segmentation algorithm, only aim to capture obvious palm vein features without collecting palm print interference features. Therefore, they do not consider different layers and thus have to discard or lose subtle features, such as capillary features and small branch veins. However, even so, some palm print interference features with relatively obvious lines will still be present.
[0063] Thus, in step S3, palmprint interference information in palm vein information is removed based on palmprint information. Since both palm vein information and palmprint information are obtained through a bidirectional connection-based segmentation algorithm, when the infrared palm vein image and the RGB palmprint image are taken from the same palm, the palmprint interference information in palm vein information will necessarily be reflected in the palmprint information. Therefore, the corresponding palmprint interference information in palm vein information can be removed based on palmprint information to obtain a result that retains both the main vein features and the features of small capillaries and branch veins.
[0064] Simultaneously, by leveraging the characteristics of different devices, irrelevant palm print and other interfering information are removed. This allows the model to focus more intently on palm vein information during training and recognition, thereby enabling more accurate differentiation of palm veins from different hands. Removing palm print interference from infrared palm vein images improves the accuracy of palm vein recognition, allowing for more precise palm vein identification.
[0065] The palm vein enhancement method based on bidirectional connectivity segmentation algorithm provided in this application, compared with the prior art, utilizes the different characteristics of image acquisition devices to acquire infrared palm vein images and RGB palmprint images respectively. First, the palm vein information and palmprint information are enhanced based on the bidirectional connectivity segmentation algorithm, preserving subtle features at different levels. Then, the information obtained after segmentation and enhancement is used to verify each other. By removing the corresponding palmprint interference information in the palm vein information based on the palmprint information, the main vein features, capillary and branch vein features are preserved. In this way, the model can focus more on the effective information of palm veins during training and recognition, thereby improving the palm vein recognition accuracy.
[0066] In another embodiment of this application, please refer to [the relevant document / reference]. Figure 4 and Figure 5 In step S2, the bidirectional connection segmentation algorithm includes the following steps:
[0067] Step S21: Extract features from the input image x through convolutional layers C1, C2, and C3 to obtain features C1(x), C2(x), and C3(x), respectively.
[0068] Step S22: Perform deconvolution operations on features C2(x) and C3(x) to obtain deconvolution results D1 and D2 respectively;
[0069] Step S23: Superimpose features C1(x), C2(x), and C3(x) with the deconvolution results D1 and D2 to obtain the level outputs zx1, zx2, and zx3 corresponding to different levels.
[0070] In step S23, the calculation formulas for hierarchical output zx1, hierarchical output zx2, and hierarchical output zx3 are as follows:
[0071] zx1=C1(x)+D1,
[0072] zx² = C²(x) + D²,
[0073] zx3=C3(x).
[0074] It is understandable that the input image x can be an infrared palm vein image or an RGB palmprint image. That is, when extracting palm vein information from an infrared palm vein image using a bidirectional connection-based segmentation algorithm, the input image x can be an infrared palm vein image; similarly, when extracting palmprint information from an RGB palmprint image using a bidirectional connection-based segmentation algorithm, the input image x can be an RGB palmprint image. The segmentation algorithm used employs bidirectional connections, meaning it collects feature information from as many different levels as possible. This is because the level outputs zx1, zx2, and zx3 correspond to different features at different levels, such as veins and capillaries in the palm veins—the former being coarse and the latter fine. By setting different levels, more subtle and effective features can be extracted from the infrared palm vein image, and more subtle palmprint information corresponding to palmprint interference information can also be extracted from the RGB palmprint image.
[0075] Further, in step S22, the method for deconvolution operation includes the following steps:
[0076] Step S231, define the input image X for the deconvolution operation as:
[0077]
[0078] The convolution kernel K is:
[0079]
[0080] Step S232, output image X 、 The size is 5×5. The input image X is padded with 0s between every two elements, i.e.:
[0081]
[0082] Step S233, map the convolution kernel K to X 、 Performing a convolution operation yields the result Y of the deconvolution operation, i.e.:
[0083]
[0084] It is understandable that the input image X for the deconvolution operation is feature C2(x) or feature C3(x), and the result Y of the deconvolution operation is the deconvolution result D1 and the deconvolution result D2.
[0085] Furthermore, after step S23, the bidirectional connection segmentation algorithm further includes the following steps:
[0086] Step S241: Perform deconvolution operations on the hierarchical output zx1 and hierarchical output zx2 to obtain deconvolution operation results D3 and D4 respectively;
[0087] Step S242: Superimpose the hierarchical outputs zx1, zx2, and zx3 with the deconvolution results D3 and D4 to obtain the hierarchical outputs zy1, zy2, and zy3 corresponding to different levels.
[0088] In step S242, the calculation formulas for hierarchical output zx1, hierarchical output zx2, and hierarchical output zx3 are as follows:
[0089] zy1 = zx1,
[0090] zy2=zx2+D3,
[0091] zy3=zx3+D4.
[0092] Furthermore, after step S242, the bidirectional connection segmentation algorithm further includes the following steps:
[0093] Step S25: Scaling operation is performed on the hierarchical outputs zy1, zy2, and zy3 to obtain a standard image with consistent resolution.
[0094] It is understandable that, through step S25, the hierarchical outputs zy1, zy2, and zy3 are scaled to obtain standard images with consistent resolution, thereby unifying the image specifications so that they can be used to train the model.
[0095] In another embodiment of this application, step S3, the method for removing palm print interference information from palm vein information based on palm print information, includes:
[0096] Step S31: Obtain the intersection information of palm vein information and palm print information.
[0097] Step S32: Remove the overlapping information from the palm vein information.
[0098] It is understandable that in step S31, the intersection information of palm vein information and palm print information is obtained. The intersection information is also the palm print features in the infrared palm vein image. By removing the intersection information from the palm vein information, effective information with only palm vein features, capillary features, and other features can be obtained.
[0099] Please refer to the following: Figure 6 This application also provides a model training method, which includes the following steps:
[0100] The palm vein enhancement method based on the bidirectional connection segmentation algorithm described above was used to obtain a large number of training datasets;
[0101] The training dataset is input into the palm vein recognition CNN model for training.
[0102] Compared with existing technologies, the model training method provided in this application adopts a palm vein enhancement method based on bidirectional connection segmentation algorithm to obtain a large training dataset. This training dataset retains subtle features at different levels. By removing the corresponding palm print interference information from the palm vein information based on the palm print information, the main vein features, capillary and branch vein features are retained. This makes the model more focused on the effective information of the palm vein during training and recognition, and improves the accuracy of the palm vein recognition CNN model in palm vein recognition.
[0103] In another embodiment of this application, the palm vein recognition CNN model is a palm vein recognition CNN model based on ResNet50.
[0104] Understandably, the ResNet50 network structure, by deepening the number of network layers, can fully utilize the characteristic of the training dataset to retain subtle features at different levels, thereby improving the segmentation accuracy of the network. More jump connections can be added in the middle of the network, which can better combine the background semantic information of the image, further improving the accuracy of the palm vein recognition CNN model in palm vein recognition.
[0105] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A palm vein enhancement method based on a bidirectional connectivity segmentation algorithm, characterized in that, include: Step S1: Acquire infrared palm vein images and RGB palm print images respectively; The method for acquiring RGB palm print images is to use an RGB camera, and the method for acquiring infrared palm vein images is to use an infrared palm vein acquisition device to acquire palm vein image information. Step S2: Extract palm vein information from infrared palm vein images using a bidirectional connection-based segmentation algorithm, and extract palm print information from RGB palm print images using a bidirectional connection-based segmentation algorithm. In step S2, the bidirectional connection segmentation algorithm includes the following steps: Step S21: Extract features from the input image x through convolutional layers C1, C2, and C3 to obtain features C1(x), C2(x), and C3(x), respectively. Step S22: Perform deconvolution operations on features C2(x) and C3(x) to obtain deconvolution results D1 and D2 respectively; Step S23: Superimpose features C1(x), C2(x), and C3(x) with the deconvolution results D1 and D2 to obtain the level outputs zx1, zx2, and zx3 corresponding to different levels. The calculation formulas for hierarchical outputs zx1, zx2, and zx3 are as follows: , , ; Step S241: Perform deconvolution operations on the hierarchical output zx1 and hierarchical output zx2 to obtain deconvolution operation results D3 and D4 respectively; Step S242: Superimpose the hierarchical outputs zx1, zx2, and zx3 with the deconvolution results D3 and D4 to obtain the hierarchical outputs zy1, zy2, and zy3 corresponding to different levels. The calculation formulas for hierarchical outputs zy1, zy2, and zy3 are as follows: , , ; Step S3, remove palmprint interference information from palm vein information based on palmprint information, specifically including: Step S31: Obtain the intersection information of palm vein information and palm print information. Step S32: Remove the overlapping information from the palm vein information.
2. The palm vein enhancement method based on bidirectional connection segmentation algorithm as described in claim 1, characterized in that, In step S22, the method for deconvolution operation includes the following steps: Step S231: Define the input image X and the convolution kernel K for the deconvolution operation; Step S232, output image X 、 The input image X is 5×5, and every two elements are padded with 0s. Step S233, convert the output image X corresponding to the convolution kernel K. 、 Perform a convolution operation to obtain the result Y of the deconvolution operation.
3. The palm vein enhancement method based on bidirectional connection segmentation algorithm as described in claim 1, characterized in that, Following step S242, the bidirectional connection segmentation algorithm further includes the following steps: Step S25: Scaling operation is performed on the hierarchical outputs zy1, zy2, and zy3 to obtain a standard image with consistent resolution.
4. A model training method, characterized in that, The model training method includes the following steps: The palm vein enhancement method based on bidirectional connection segmentation algorithm as described in any one of claims 1 to 3 is used to obtain a large number of training datasets; The training dataset is input into the palm vein recognition CNN model for training.
5. The model training method as described in claim 4, characterized in that, The palm vein recognition CNN model is a palm vein recognition CNN model based on ResNet50.