A finger sweat gland extraction method based on deep compressed neural network
By directly extracting sweat gland features from OCT data through a deep compressed neural network, the complexity and time-consuming problems of existing methods are solved, efficient and robust end-to-end extraction is achieved, and high-precision surface sweat pore images are generated.
Patent Information
- Application Number
- CN202411153540.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-21
- Publication Date
- 2025-09-23
- Estimated Expiration
- 2044-08-21
AI Technical Summary
Existing neural network-based finger sweat gland extraction methods are complex, time-consuming, and not robust to noise. They cannot directly extract sweat glands from raw OCT data and require additional mapping steps and denoising operations.
A deep compression neural network is used to extract sweat gland features directly from raw OCT data through an expanded deep compression module and a basic deep compression module combined with a three-dimensional convolutional layer, including three-dimensional expanded convolution, random dropout layer and binarization processing, which simplifies the process and improves robustness.
It achieves end-to-end extraction from three-dimensional volume data to two-dimensional images, shortens extraction time, improves extraction accuracy and speed, enhances robustness to noise, and generates accurate surface sweat pore images directly from OCT raw data.
Smart Images
Figure CN119027988B_ABST
Abstract
Description
Technical Field
[0001] The present application belongs to the field of fingerprint recognition technology, and in particular relates to a finger sweat gland extraction method based on a deep compressed neural network. Background Art
[0002] Due to their uniqueness, durability, and ease of capture, fingerprints, as the most commonly used biometric features, are widely used for personal identification. Fingerprint features are generally classified into three levels, with fingerprint collection systems typically collecting primary and secondary fingerprint features for identification. With the advancement of collection equipment, tertiary features are gaining importance and application. Among these, surface sweat pores have been proven to be a discriminant feature with high personal identification capabilities. However, the surface of a finger can age and wear due to age and other environmental factors, thus damaging the fingerprint and rendering it incapable of identification. Furthermore, forged fingerprint samples made of materials such as silicone can often deceive these systems.
[0003] Biological evidence suggests that the sweat pores on the surface of a finger are the surface openings of the internal sweat glands. This suggests that it is possible to obtain highly resistant and anti-counterfeit sweat gland features from within the finger, thereby obtaining corresponding surface sweat pore information. Furthermore, optical coherence tomography (OCT), a non-invasive imaging technology, can acquire information 1 to 3 mm below the surface of human skin, generating 3D volumetric data of fingerprints. This makes it possible to acquire high-resolution, three-dimensional images of internal sweat glands.
[0004] Existing neural network-based methods for extracting sweat glands from fingertips using OCT are generally based on volume data or volume data slices, finding the location of sweat glands in the OCT data and mapping them into surface sweat pores. These methods require preprocessing of the speckle noise present in the OCT data and require additional mapping steps to obtain the corresponding surface sweat pore images. Neural networks use U-shaped or improved U-shaped structures for segmentation, which is time-consuming and prone to losing small sweat gland features during the downsampling process. Furthermore, existing methods are not robust to noise and cannot directly extract sweat glands from raw OCT data. De-noising is usually required first, which is complex and time-consuming.
[0005] Currently, fingerprint recognition utilizes sweat pore information. The existing neural network-based finger sweat gland extraction method does not achieve end-to-end extraction from sweat glands to sweat pores, requires additional mapping steps, and the process is complicated. Summary of the Invention
[0006] The purpose of this application is to provide a finger sweat gland extraction method based on deep compressed neural network to overcome the problems of existing OCT finger sweat gland extraction, such as complex process, long time consumption and weak noise robustness.
[0007] In order to achieve the above objectives, the technical solutions of this application are as follows:
[0008] A finger sweat gland extraction method based on deep compression neural network, comprising:
[0009] Segment the OCT fingerprint data and input the resulting series of sub-blocks into a deep compression neural network;
[0010] After continuous expansion and depth compression modules, the dilated convolution is used to improve the receptive field of the sweat glands and extract the preliminary features of the sweat glands;
[0011] After random dropout, the sweat gland preliminary features are extracted through a continuous basic deep compression module to extract the sweat gland refined features;
[0012] After random inactivation, the sweat gland refined features pass through successive convolutional layers to output the sweat pore feature map;
[0013] The sweat pore feature map is binarized to obtain the surface sweat pore image, and the surface sweat pore images corresponding to all sub-blocks are spliced to obtain the finger sweat pore map.
[0014] Furthermore, the dilated depth compression module includes a three-dimensional dilated convolution layer, two three-dimensional convolution layers and a three-dimensional maximum pooling layer;
[0015] The basic depth compression module includes three three-dimensional convolutional layers and one three-dimensional maximum pooling layer.
[0016] Furthermore, the pooling kernel size of the maximum pooling layer in the basic depth compression module is smaller than the pooling kernel size of the maximum pooling layer in the expanded depth compression module.
[0017] Furthermore, the deep compression neural network includes three expanded deep compression modules and two basic deep compression modules.
[0018] Furthermore, the pooling kernel sizes of the maximum pooling layers in the three expanded deep compression modules are 4×1×1, 4×1×1 and 5×1×1 respectively; the pooling kernel sizes of the maximum pooling layers in the two basic deep compression modules are 3×1×1 and 2×1×1 respectively.
[0019] Furthermore, the deep compression neural network uses binary cross entropy loss during training.
[0020] Compared with the existing technology, the beneficial effects of this application are as follows: the robustness of OCT sweat gland extraction is improved through deep compression neural networks, and OCT raw data can be directly extracted; at the same time, compared with ordinary neural network extraction methods, this application realizes end-to-end extraction from three-dimensional volume data to two-dimensional images, simplifies the process, and greatly shortens the extraction time.
[0021] This application proposes a finger sweat gland extraction method based on a deep compression neural network. A deep compression module is designed to address the narrow and elongated morphology of sweat glands. This characteristic of sweat glands allows them to display richer and more continuous feature information at depth. Compression at depth maximizes the condensation of sweat gland features while providing robustness to noise, enabling direct processing of raw OCT data. By extracting sweat gland features using the deep compression module, sweat glands are extracted directly from the raw OCT data without the need for denoising or mapping operations. This compression generates accurate surface sweat pore images, improving extraction accuracy and speed. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] Figure 1 This is a flow chart of the finger sweat gland extraction method based on deep compressed neural network in this application.
[0023] Figure 2 This is the structure diagram of the deep compression neural network for this application.
[0024] Figure 3 This is a schematic diagram of the structure of the expanded deep compression module according to an embodiment of the present application.
[0025] Figure 4 This is a schematic diagram of the basic deep compression module structure of an embodiment of this application. DETAILED DESCRIPTION
[0026] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.
[0027] One embodiment of the present application, such as Figure 1 As shown in the figure, a finger sweat gland extraction method based on deep compressed neural network is proposed, including:
[0028] Step S1: Segment the OCT fingerprint data, and input the resulting series of sub-blocks into a deep compression neural network.
[0029] In order to adapt to lightweight applications and reduce the memory requirements of the device, this application divides the large-sized optical coherence tomography (OCT) fingerprint volume data into a series of small-sized sub-blocks.
[0030] For example, the original size of the OCT volume data to be predicted is 480×1800×1400, which is segmented into 480×120×1400 sizes to obtain a series of sub-block data, each of which has a size of 480×120×1400.
[0031] This embodiment constructs a deep compression neural network model and trains it, and uses the trained deep compression neural network model to extract the finger sweat pore map.
[0032] To train a deep compression neural network model, we first need to obtain training samples. The OCT fingerprint volume data is H × W × N in size, consisting of N OCT images with a resolution of H × W, representing vertical slices of N consecutive fingerprints. Complete OCT volume data images are selected, and the sweat gland regions are manually annotated. The sweat glands are then mapped vertically onto the surface to obtain an annotated image of the surface sweat pores corresponding to the OCT volume data, forming the annotated dataset required for neural network model training.
[0033] To obtain training samples, we first manually annotate several sets of complete OCT volume data images and mark the sweat gland areas in the images; then, based on the annotated sweat gland areas in the images, we vertically map them to the surface to generate surface sweat pore annotated images; finally, we horizontally invert the OCT ROI volume data and its corresponding annotation images to obtain more training samples.
[0034] Step S2: After continuous expansion and depth compression modules, the receptive field of the sweat glands is improved by using the expansion convolution to extract the preliminary features of the sweat glands.
[0035] In a specific embodiment, the constructed deep compression neural network model is as follows: Figure 2 As shown in the figure, each of the first to third parts is an expansion deep compression module; each of the fourth and seventh parts is a three-dimensional random inactivation layer (dropout layer) with a dropout probability of 0.5, indicating that there is a 50% probability of setting the neuron output to 0; each of the fifth and sixth parts is a basic deep compression module; each of the eighth to tenth parts is a three-dimensional convolution layer with a convolution kernel of 3×3×3, which ultimately outputs the sweat pore feature image.
[0036] It should be noted that the number of expanded deep compression modules, basic deep compression modules and three-dimensional convolutional layers can be adjusted according to the actual application of the network. Those skilled in the art can make settings based on actual applications, which will not be repeated here.
[0037] In a specific embodiment, Figure 3 As shown in the figure, the dilated deep compression module consists of a 3D dilated convolutional layer, two 3D convolutional layers, and a 3D max pooling layer. The 3D dilated convolutional layer is processed by a 1×2×2 dilated kernel and a RELU activation function; each 3D convolutional layer is processed by a 3×3×3 convolutional kernel and a RELU activation function.
[0038] In the three dilated depth compression modules, each 3D maximum pooling layer passes through a pooling kernel of size 4×1×1, 4×1×1, and 5×1×1, respectively.
[0039] In this embodiment, the dilated depth compression module introduces dilated convolutions based on the basic depth compression module. This effectively increases the receptive field of the sweat glands, amplifies their features, and improves extraction accuracy. Furthermore, this module uses a larger pooling kernel to rapidly aggregate depth-dependent feature information of the sweat glands, speeding up sweat gland extraction and enhancing robustness to noise.
[0040] Specifically, the 3D dilated convolution layer increases the receptive field of the convolution kernel without increasing the number of parameters or computational complexity, thereby improving the accuracy and efficiency of sweat gland feature extraction. The 3D convolution layer can extract sweat gland features from 3D data. The 3D max pooling layer selects the maximum value in the pooling window as the output, enhancing sweat gland features and reducing noise interference. It also compresses information in the depth dimension, achieving depth compression.
[0041] In this step, the preliminary features of the sweat glands are obtained through continuous expansion and depth compression modules.
[0042] Step S3: After random inactivation of the sweat gland preliminary features, the sweat gland refined features are extracted through a continuous basic deep compression module.
[0043] The three-dimensional dropout layer is a regularization technique widely used in deep learning to prevent overfitting during neural network training. This example applies dropout to the preliminary and refined sweat gland features extracted by the expanded and basic deep compression modules, respectively, to prevent overfitting and improve accuracy.
[0044] In a specific embodiment, the basic deep compression module, such as Figure 4 As shown in the figure, it includes three 3D convolutional layers and one 3D max pooling layer. Each 3D convolutional layer is processed by a 3×3×3 convolution kernel and a RELU activation function. For the two basic deep compression modules, the 3D max pooling layer is processed by a 3×1×1 and 2×1×1 pooling kernel, respectively.
[0045] It is worth noting that compared with the expanded depth compression module, the pooling kernel size of the maximum pooling layer in the basic depth compression module of this embodiment is smaller than the pooling kernel size of the maximum pooling layer in the expanded depth compression module, that is, the maximum pooling kernel of the basic depth compression module uses a smaller pooling kernel, which can further refine the sweat gland features that have been obtained and improve the extraction accuracy.
[0046] Specifically, the 3D convolutional layer in the basic depth compression module extracts sweat gland features from 3D data. The 3D max pooling layer selects the maximum value within the pooling window as its output, enhancing sweat gland features and reducing noise interference while compressing the depth dimension to achieve depth compression. Through successive basic depth compression modules, refined sweat gland features are obtained.
[0047] Step S4: After random inactivation, the sweat gland refined features pass through successive convolutional layers to output the sweat pore feature map.
[0048] Specifically, each of the eighth to tenth parts is a three-dimensional convolution layer with a convolution kernel of 3×3×3. Three consecutive three-dimensional convolution layers can further extract the deep features of the sweat glands, capture more subtle boundary and texture information, and obtain the final sweat pore feature map.
[0049] Step S5: binarize the sweat pore feature map to obtain a surface sweat pore image, and splice the surface sweat pore images corresponding to all sub-blocks to obtain a finger sweat pore map.
[0050] Specifically, each 480×120×1400 sub-block is processed through a deep compression neural network model and then binarized to generate a 120×1400 surface pore image, showing white pores against a black background. To obtain a complete pore image, the sub-blocks are sequentially spliced along the segmentation direction to obtain a 1800×1400 pore map, which is the fingerprint pore image used for identification.
[0051] In a specific embodiment, a deep compressed neural network model is constructed, training parameters and loss functions are set, and the deep compressed neural network model is trained using a labeled data set to obtain a trained deep compressed neural network model, including the following steps:
[0052] Since the spatial morphology of sweat glands in OCT volume data is long, narrow, small and discontinuous, and the features are concentrated in the depth dimension, a deep compression neural network is constructed to deeply compress the OCT volume data, thereby realizing the conversion from three-dimensional OCT volume data to two-dimensional images. The input OCT volume data size is 480×120×1400, and the output image size is 120×1400.
[0053] Determine the parameters of the deep compression neural network, set the batch size to 1, and load the training OCT volume data into a batch of 480×120×1400 blocks into the deep compression neural network model for training. The number of iterations is 900 to obtain the trained network.
[0054] The parameter updates of each network layer are calculated using Adaptive Moment Estimation (Adam) with a learning rate, where the learning rate is set to 0.001.
[0055] Using binary cross entropy loss, its function form is as follows:
[0056]
[0057] In the above formula, represents the true label, is the predicted probability.
[0058] Finally, the trained deep compression neural network model predicts the surface pore images corresponding to the unlabeled OCT volume data in blocks. To accommodate the input volume data size of the trained deep compression neural network, the OCT volume data to be predicted is segmented into a series of sub-blocks of 480×120×1400. These sub-blocks are then fed into the trained deep compression neural network to obtain the corresponding 120×1400 surface pore sub-images. The sub-images are then stitched together to form a complete surface pore image.
[0059] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.
Claims
1. A finger sweat gland extraction method based on deep compressed neural network, characterized in that: The finger sweat gland extraction method based on deep compression neural network includes: Segment the OCT fingerprint data and input the resulting series of sub-blocks into a deep compression neural network; After continuous expansion and depth compression modules, the dilated convolution is used to improve the receptive field of the sweat glands and extract the preliminary features of the sweat glands; After random dropout, the sweat gland preliminary features are extracted through a continuous basic deep compression module to extract the sweat gland refined features; After random inactivation, the sweat gland refined features pass through successive convolutional layers to output the sweat pore feature map; The sweat pore feature map is binarized to obtain the surface sweat pore image, and the surface sweat pore images corresponding to all sub-blocks are spliced to obtain the finger sweat pore map.
2. The finger sweat gland extraction method based on deep compression neural network according to claim 1, characterized in that: The dilated depth compression module includes a three-dimensional dilated convolution layer, two three-dimensional convolution layers and a three-dimensional maximum pooling layer; The basic depth compression module includes three three-dimensional convolutional layers and one three-dimensional maximum pooling layer.
3. The finger sweat gland extraction method based on deep compression neural network according to claim 2, characterized in that: The pooling kernel size of the maximum pooling layer in the basic depth compression module is smaller than the pooling kernel size of the maximum pooling layer in the expanded depth compression module.
4. The finger sweat gland extraction method based on deep compression neural network according to claim 3, characterized in that: The deep compression neural network includes three expanded deep compression modules and two basic deep compression modules.
5. The finger sweat gland extraction method based on deep compression neural network according to claim 4, characterized in that: The pooling kernel sizes of the maximum pooling layers in the three expanded deep compression modules are 4×1×1, 4×1×1 and 5×1×1 respectively; the pooling kernel sizes of the maximum pooling layers in the two basic deep compression modules are 3×1×1 and 2×1×1 respectively.
6. The finger sweat gland extraction method based on deep compression neural network according to claim 1, characterized in that: The deep compression neural network uses binary cross entropy loss during training.