WO2021056972A1 - 基于神经网络和概率图模型的手指静脉分割方法及装置 - Google Patents

基于神经网络和概率图模型的手指静脉分割方法及装置 Download PDF

Info

Publication number
WO2021056972A1
WO2021056972A1 PCT/CN2020/078177 CN2020078177W WO2021056972A1 WO 2021056972 A1 WO2021056972 A1 WO 2021056972A1 CN 2020078177 W CN2020078177 W CN 2020078177W WO 2021056972 A1 WO2021056972 A1 WO 2021056972A1
Authority
WO
WIPO (PCT)
Prior art keywords
neural network
finger vein
training
graph model
probability graph
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2020/078177
Other languages
English (en)
French (fr)
Inventor
曾军英
王璠
秦传波
朱伯远
朱京明
翟懿奎
甘俊英
邓建祥
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Wuyi University Fujian
Original Assignee
Wuyi University Fujian
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Wuyi University Fujian filed Critical Wuyi University Fujian
Publication of WO2021056972A1 publication Critical patent/WO2021056972A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
    • G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/14—Vascular patterns
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Definitions

  • the invention relates to the technical field of neural networks, in particular to a finger vein segmentation method and device based on neural networks and probability graph models.
  • Finger vein recognition as one of the many biometric recognition technologies, has become a hot spot of current research due to its advantages of non-contact collection, live detection, not easy to forge, and low cost.
  • the segmentation of finger vein images is a key step in the vein recognition technology. The quality of the segmentation effect directly affects the accuracy and accuracy of subsequent recognition.
  • the traditional CNN has a convolution filter with a larger receptive field, and the pixel-level label generated after reconstruction will produce a coarse output (in the last layer of the network, each neuron corresponds to a large area of the original picture).
  • the max pooling layer in CNNs further reduces the chance of obtaining fine segmentation output. For example, in semantic segmentation tasks, this may result in unclear boundaries and drop-like shapes.
  • CNN lacks smoothing constraints that encourage label consistency between similar pixels, as well as the spatial and appearance consistency of label output. The lack of such smoothing constraints will result in poor target contours and smaller false regions in the segmentation output.
  • CRFs Markov Random Fields
  • CRFs Conditional Random Fields
  • the core idea of semantically labeled CRF reasoning is to express the label assignment problem as a probabilistic reasoning problem, which includes assumptions such as label consistency between similar pixels. CRF inference can refine weak and rough pixel-level label predictions, resulting in clear boundaries and fine-grained segmentation. Therefore, intuitively, CRFs can be used to overcome the shortcomings of using CNN for pixel-level labeling tasks.
  • the present invention aims to solve at least one of the technical problems existing in the prior art.
  • the present invention proposes a finger vein segmentation method and device based on neural network and probability graph model, which can use dense conditional random field (CRF) with Gaussian paired potential as a recurrent neural network (RNN) to establish an average Field approximate reasoning, it can refine the coarse output of the traditional neural network (CNN) in the forward pass process, and at the same time pass the error back to the neural network (CNN) during the training process. After a certain period of training, due to the full use of the dense The advantage of Conditional Random Field (CRF), so it can produce finer vein segmentation.
  • CCF dense conditional random field
  • RNN recurrent neural network
  • a finger vein segmentation method based on neural network and probability graph model includes the following steps: generating a preprocessed image that is conducive to neural network training and testing; generating a gold standard that is conducive to neural network training; The dense conditional random field of the Gaussian paired potential is used as a cyclic neural network for average field approximate reasoning; the dense conditional random field is integrated into the neural network, and the neural network training processing is performed.
  • a finger vein segmentation method based on a neural network and a probability graph model has at least the following beneficial effects:
  • the finger vein segmentation method based on a neural network and a probability graph model of the present invention can combine Dense Conditional Random Field (CRF) with Gaussian paired potential as a kind of Recurrent Neural Network (RNN) establishes the average field approximate reasoning, which can refine the coarse output of the traditional neural network (CNN) in the forward pass process, and at the same time In the training process, the error is passed back to the neural network (CNN). After a certain period of training, due to the full use of the advantages of the dense conditional random field (CRF), a finer vein segmentation can be generated.
  • CCF Dense Conditional Random Field
  • RNN Recurrent Neural Network
  • the generating of a preprocessed image that is beneficial for neural network training and testing includes: the preprocessing process of the image includes grayscale transformation processing, ROI extraction processing, normalization processing, and adaptive histogram with limited contrast. Picture equalization processing and gamma adjustment processing.
  • the generation of the gold standard for neural network training includes: extracting six finger veins through repeated line tracking, wide line detector, Gabor filter, maximum curvature, average curvature, and enhanced maximum curvature Divide the ridge map, and attach the six ridge maps with different weights to obtain the golden standard for neural network training.
  • using the dense conditional random field of Gaussian paired potentials as a cyclic neural network to perform average field approximate inference includes: the neural network decomposition process of the dense conditional random field includes initialization, message passing, and weighted filtering.
  • a finger vein segmentation device based on a neural network and a probability graph model includes: a preprocessed image generating unit for generating preprocessed images that are beneficial to neural network training and testing; gold standard generation The unit is used to generate the gold standard that is conducive to neural network training; the average field approximate inference unit is used to use the dense conditional random field of Gaussian paired potentials as a recurrent neural network for average field approximate inference; the training processing unit is used to The dense conditional random field is integrated into the neural network for training and processing of the neural network.
  • a finger vein segmentation device based on a neural network and a probability graph model includes: at least one control processor and a memory for communicating with the at least one control processor; said The memory stores instructions executable by the at least one control processor, and the instructions are executed by the at least one control processor, so that the at least one control processor can execute the neural-based Finger vein segmentation method based on network and probability graph model.
  • a computer-readable storage medium according to an embodiment of the fourth aspect of the present invention, wherein the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to make a computer execute the above-mentioned embodiment of the first aspect Finger vein segmentation method based on neural network and probability graph model.
  • a computer program product includes a computer program stored on a computer-readable storage medium, the computer program includes program instructions, when the program instructions are executed by a computer At the time, the computer is caused to execute the finger vein segmentation method based on the neural network and the probability graph model as in the above-mentioned first aspect embodiment.
  • FIG. 1 is a flowchart of steps of a method for segmenting finger veins based on neural network and probability graph model according to an embodiment of the present invention
  • FIG. 2 is a framework flowchart of a finger vein segmentation method based on neural network and probability graph model according to an embodiment of the present invention
  • FIG. 3 is a flowchart of a method for segmenting finger veins based on a neural network and a probability graph model to generate a gold standard for neural network training according to an embodiment of the present invention
  • FIG. 4 is a schematic diagram of a common CNN operation process of decomposing the average field in a dense conditional random field in a finger vein segmentation method based on a neural network and a probability graph model according to an embodiment of the present invention
  • FIG. 5 is a schematic diagram of an iterative mean field approximating a recurrent neural network in a finger vein segmentation method based on a neural network and a probability graph model according to an embodiment of the present invention
  • FIG. 6 is a flow chart of a common CNN operation algorithm that decomposes the average field in the dense CRF into a common CNN operation algorithm in a finger vein segmentation method based on a neural network and a probability graph model according to an embodiment of the present invention
  • FIG. 7 is a structural diagram of a finger vein segmentation device based on neural network and probability graph model according to an embodiment of the present invention.
  • FIG. 8 is a structural diagram of a finger vein segmentation device based on a neural network and a probability graph model according to an embodiment of the present invention.
  • Finger vein segmentation device 100 based on neural network and probability graph model, preprocessed image generating unit 110, gold standard generating unit 120, mean field approximate reasoning unit 130, training processing unit 140;
  • a finger vein segmentation device 200 based on a neural network and a probability graph model.
  • orientation description involved such as up, down, front, back, left, right, etc. indicates the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, but In order to facilitate the description of the present invention and simplify the description, it does not indicate or imply that the device or element referred to must have a specific orientation, be configured and operate in a specific orientation, and therefore cannot be understood as a limitation to the present invention.
  • a method for segmenting finger veins based on a neural network and a probability graph model includes the following steps:
  • S1 Generate preprocessed images that are conducive to neural network training and testing
  • S2 Generate a gold standard that is conducive to neural network training; the gold standard is a label;
  • S4 Integrate the dense conditional random field into the neural network, and perform training processing of the neural network.
  • a finger vein segmentation method based on a neural network and a probability graph model has at least the following beneficial effects:
  • the finger vein segmentation method based on a neural network and a probability graph model of the present invention can combine Dense Conditional Random Field (CRF) with Gaussian paired potential as a kind of Recurrent Neural Network (RNN) establishes the average field approximate reasoning, which can refine the coarse output of the traditional neural network (CNN) in the forward pass process, and at the same time In the training process, the error is passed back to the neural network (CNN). After a certain period of training, due to the full use of the advantages of the dense conditional random field (CRF), a finer vein segmentation can be generated.
  • CCF Dense Conditional Random Field
  • RNN Recurrent Neural Network
  • the generating of a preprocessed image that is beneficial for neural network training and testing includes: the preprocessing process of the image includes grayscale transformation processing, ROI extraction processing, normalization processing, and adaptive histogram with limited contrast. Picture equalization processing and gamma adjustment processing. The purpose is to highlight areas with rich vein patterns, reduce processing time and increase accuracy.
  • the generation of the gold standard for neural network training includes: extracting six finger veins through repeated line tracking, wide line detector, Gabor filter, maximum curvature, average curvature, and enhanced maximum curvature Divide the ridge map, and attach the six ridge maps with different weights to obtain the golden standard for neural network training.
  • This method can give full play to the advantages of each traditional extraction texture, make up for the disadvantages, and make the label as accurate as possible.
  • using the dense conditional random field of Gaussian paired potentials as a cyclic neural network to perform average field approximate inference includes: the neural network decomposition process of the dense conditional random field includes initialization, message passing, and weighted filtering.
  • the traditional CNN has a convolution filter with a larger receptive field.
  • the pixel-level label generated after reconstruction will produce a coarse output, and there is no reference standard suitable for neural network training, etc.
  • Fig. 2 is a flow chart of the overall framework of a finger vein segmentation method according to an embodiment of the present invention, including four parts: image preprocessing, average field approximate reasoning, label making, and network building.
  • Image preprocessing includes grayscale transformation, ROI extraction, normalization, adaptive histogram equalization with limited contrast, and gamma adjustment. Since the background contrast of single-channel veins is better than RGB images, the original RGB vein images are converted into single-channel images, and then normalized and contrast-limited adaptive histogram equalization is used to enhance the front and back background contrast, and finally gamma is introduced Adjust to further improve the image quality.
  • Figure 3 is the production process of the gold standard.
  • six traditional methods are selected as the basic structure segmentation diagram of the gold standard (ie label). According to the quality of the segmentation and the pros and cons of various methods, different weights are attached. The size of is marked in Figure 3, and then it is subjected to image superposition, averaging, binarization, filtering and other operations to obtain the final gold standard for network training.
  • Figure 4 is the decomposition of the average field in a dense random field into a common CNN operation process.
  • Figure 6 is a flow chart of the average field decomposition in dense CRF into a common CNN operation algorithm, including initialization, message passing, weighted filter output, compatibility transformation, adding unary potential and normalization steps. This step is described in detail below.
  • Q(l) is a probability picture, corresponding to the probability that each pixel of the original image belongs to category l.
  • k(f i , f j ) is a Gaussian kernel weight, which is equivalent to the convolution operation of a neural network.
  • the specific k(f i , f j ) generally includes location, color measurement, etc. as features.
  • message passing is achieved by applying an M Gaussian filter to the Q value.
  • Gaussian filter coefficients are obtained based on image characteristics such as pixel position and RGB value. These characteristics reflect how strong the relationship between a pixel and other pixels is.
  • the receptive field of each filter spans the entire image, so it is impossible to rely on brute force to implement the filter.
  • the error derivative to the filter output is sent back using the same M Gaussian filter to calculate the error derivative to the filter input.
  • this can be achieved by only reversing the order of the separable filters in the blur phase, while constructing the Permutohedral lattice, splatting and slicing in the same way as the forward channel. Therefore, the back propagation through this filtering stage can also be performed in O(N) time.
  • This step can be seen as using a 1*1 convolutional layer to perform convolutional layer operations on multiple feature maps. After the calculation, a new probability map is output for every two feature maps.
  • the next step of the mean field iteration is to take the weighted sum of the M filter outputs (each class label l) of the previous step.
  • each class label is considered individually, this can be seen as a usual convolution with a 1 ⁇ 1 filter with M input channels and one output channel. Since the input and output of this step are known during back propagation, the error derivative with respect to the filter weight can be calculated, which makes it possible to automatically learn the filter weight (each Gaussian in the previous stage) The relative contribution of the filter output).
  • the error derivative of the input can also be calculated in the usual way, so as to pass the error derivative to the previous stage. In order to obtain more adjustable parameters, we use independent kernel weights for each class label.
  • the output of the previous step is shared to varying degrees (in Algorithm 1 by Said).
  • the compatibility between the two labels l and l' is parameterized by the label compatibility function ⁇ (l, l').
  • Potts model, by ⁇ (l,l′) [l ⁇ l′], where [ ⁇ ] is the Iverson bracket (the condition in the square bracket is satisfied, then it is 1, if it is not satisfied, it is 0), if it is similar Pixels of attributes are assigned different labels, and fixed penalties are assigned.
  • the limitation of this model is that it assigns the same penalty to all different tag pairs. Intuitively, better results can be obtained by considering the compatibility between different tag pairs and penalizing allocation accordingly.
  • the compatibility transformation step can be regarded as another convolutional layer, in which the spatial receptive field of the filter is 1 ⁇ 1, and the number of input and output channels are both L. Learning the weight of the filter is equivalent to learning the label compatibility function ⁇ . Since this step is a common convolution operation, the error differential can be transferred from the output of this step to the input.
  • the output of the compatibility transformation stage subtracts the corresponding element from the unary input U.
  • the error can be easily transmitted by copying the difference at the output of this step to two inputs with appropriate signs.
  • the iterative normalization step can be regarded as another softmax operation without parameters.
  • the differential at the output of this step can be passed to the input using the backpropagation of the softmax operation.
  • Figure 5 is a schematic diagram of approximating the iterative average field to a cyclic neural network. The following describes the implementation process in detail.
  • the entire image (or multiple images) can be used as a small batch processing, and an appropriate loss function, such as the softmax loss related to the label segmentation of the image, can be used to calculate the error of each pixel output in the network.
  • an appropriate loss function such as the softmax loss related to the label segmentation of the image
  • a finger vein segmentation device 100 based on a neural network and a probability graph model includes:
  • the preprocessed image generating unit 110 is used to generate preprocessed images that are beneficial to neural network training and testing;
  • the gold standard generating unit 120 is used to generate a gold standard that is conducive to neural network training
  • the average field approximate reasoning unit 130 is configured to use the dense conditional random field of the Gaussian paired potential as a cyclic neural network to perform average field approximate reasoning;
  • the training processing unit 140 is used to integrate the dense conditional random field into the neural network to perform neural network training processing.
  • the finger vein segmentation device 100 based on neural network and probability graph model in this embodiment is based on the same inventive concept as the above-mentioned finger vein segmentation method based on neural network and probability graph model, the method embodiment The corresponding content in is also applicable to this device embodiment, and will not be described in detail here.
  • a finger vein segmentation device 200 based on a neural network and a probability graph model can be any type of intelligence. Terminals, such as mobile phones, tablet computers, personal computers, etc.
  • the finger vein segmentation device 200 based on a neural network and a probability graph model includes: one or more control processors 210 and a memory 220.
  • one control processor 210 is taken as an example.
  • control processor 210 and the memory 220 may be connected through a bus or in other ways. In FIG. 8, the connection through a bus is taken as an example.
  • the memory 220 can be used to store non-transitory software programs, non-transitory computer-executable programs and modules, such as the finger based on neural network and probability graph model in the embodiment of the present invention.
  • the program instructions/modules corresponding to the vein segmentation method are, for example, the units 110-140 shown in FIG. 7.
  • the control processor 210 executes various functional applications and data processing of the finger vein segmentation device 100 based on neural networks and probability graph models by running non-transient software programs, instructions, and modules stored in the memory 220, thereby realizing the above method The finger vein segmentation method based on neural network and probability graph model of the embodiment.
  • the memory 220 may include a storage program area and a storage data area.
  • the storage program area may store an operating system and an application program required by at least one function; the storage data area may store the finger vein segmentation device 100 based on neural networks and probability graph models. Use the created data, etc.
  • the memory 220 may include a high-speed random access memory, and may also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state storage devices.
  • the memory 220 may optionally include a memory 220 remotely provided with respect to the control processor 210, and these remote memories 220 may be connected to the finger vein segmentation device 200 based on a neural network and a probability graph model through a network. Examples of the aforementioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.
  • the one or more modules are stored in the memory 220, and when executed by the one or more control processors 210, the finger vein segmentation method based on the neural network and the probability graph model in the above method embodiment is executed, For example, the method steps S1 to S4 in FIG. 1 described above are executed to realize the functions of the units 110-140 in FIG. 7.
  • a computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are executed by one or more control processors 210, for example, Executed by one control processor 210 in FIG. 8 can cause the above-mentioned one or more control processors 210 to execute the finger vein segmentation method based on the neural network and the probability graph model in the above-mentioned method embodiment, for example, execute the above-described diagram
  • the method steps S1 to S4 in 1 realize the functions of the units 110-140 in FIG. 7.
  • the device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the objectives of the solutions of the embodiments.
  • each implementation manner can be implemented by means of software plus a general hardware platform.
  • All or part of the processes in the methods of the above embodiments can be implemented by computer programs instructing relevant hardware.
  • the programs can be stored in a computer readable storage medium. At this time, it may include the flow of the embodiment of the above-mentioned method.
  • the storage medium may be a magnetic disk, an optical disc, a read-only memory (Read Only Memory, ROM), or a random access memory (Random Access Memory, RAM), etc.
  • a computer program product includes a computer program stored on a computer-readable storage medium, the computer program includes program instructions, when the program instructions are executed by a computer At this time, the computer is made to execute the finger vein segmentation method based on the neural network and the probability graph model as in the embodiment of the first aspect described above.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Multimedia (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Human Computer Interaction (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
  • Image Analysis (AREA)

Abstract

一种基于神经网络和概率图模型的手指静脉分割方法及装置，包括以下步骤：生成利于神经网络训练和测试的预处理图像(S1)；生成利于神经网络训练的金标准(S2)；将高斯成对势的密集条件随机场作为循环神经网络进行平均场近似推理(S3)；将所述密集条件随机场融入至神经网络中，进行神经网络的训练处理(S4)。上述方案能够将具有高斯成对势的密集条件随机场(CRF)作为一种循环神经网络(RNN)建立了平均场近似推理，它可以在前向传递过程中细化传统神经网络(CNN)的粗输出，同时在训练过程中将误差传递回神经网络(CNN)，经过一定时间的训练，由于充分利用了密集条件随机场(CRF)的优势，因此可以产生更精细的静脉分割。

Description

基于神经网络和概率图模型的手指静脉分割方法及装置 技术领域
本发明涉及神经网络技术领域，特别涉及一种基于神经网络和概率图模型的手指静脉分割方法及装置。
背景技术
近年来，随着人们对生物识别系统安全性和准确性的要求越来越高，生物特征识别技术得到越来越多的关注。手指静脉识别作为众多生物特征识别技术中的一种，由于具有非接触式采集、活体检测、不易伪造、成本较低等优点，成为当前研究的热点。而手指静脉图像血管的分割是指静脉识别技术中的一个关键步骤，分割效果的优劣直接影响后续识别的精度和准确度。
传统的CNN具有感受野较大的卷积滤波器，重构后生成像素级标签会产生粗输出(在网络最后一层的时候，每个神经元对应到原始图片的一块很大区域)。在CNNs中存在的max pooling层进一步降低了获得精细分割输出的机会。例如，在语义分割任务中，这可能导致边界不清晰和水滴状形状。其次，CNN缺乏鼓励相似像素之间的标签一致性，以及标签输出的空间和外观一致性的平滑约束，缺乏这样的平滑约束，会导致分割输出中目标轮廓较差，伪区域较小。
随着深度学习技术的发展，概率图模型已成为提高像素级标记任务精度的有效方法。特别是马尔可夫随机场(MRFs)及其变体条件随机场(CRFs)在这一领域获得了广泛的成功，并已成为计算机视觉中最成功的图模型之一。语义标注的CRF推理的核心思想是将标签分配问题表示为一个概率推理问题，其中包含了相似像素之间的标签一致性等假设。CRF推理能够细化弱和粗糙的像素级标签预测，从而产生清晰的边界和细粒度的分割。因此，直观地，CRFs可以用来克服利用CNN进行像素级标记任务的缺点。
发明内容
本发明旨在至少解决现有技术中存在的技术问题之一。为此，本发明提出一 种基于神经网络和概率图模型的手指静脉分割方法及装置，能够将具有高斯成对势的密集条件随机场(CRF)作为一种循环神经网络(RNN)建立了平均场近似推理，它可以在前向传递过程中细化传统神经网络(CNN)的粗输出，同时在训练过程中将误差传递回神经网络(CNN)，经过一定时间的训练，由于充分利用了密集条件随机场(CRF)的优势，因此可以产生更精细的静脉分割。
根据本发明的第一方面实施例的一种基于神经网络和概率图模型的手指静脉分割方法，包括以下步骤：生成利于神经网络训练和测试的预处理图像；生成利于神经网络训练的金标准；将高斯成对势的密集条件随机场作为循环神经网络进行平均场近似推理；将所述密集条件随机场融入至神经网络中，进行神经网络的训练处理。
根据本发明第一方面实施例的一种基于神经网络和概率图模型的手指静脉分割方法，至少具有如下有益效果：本发明的一种基于神经网络和概率图模型的手指静脉分割方法，能够将具有高斯成对势的密集条件随机场(CRF)作为一种循环神经网络(RNN)建立了平均场近似推理，它可以在前向传递过程中细化传统神经网络(CNN)的粗输出，同时在训练过程中将误差传递回神经网络(CNN)，经过一定时间的训练，由于充分利用了密集条件随机场(CRF)的优势，因此可以产生更精细的静脉分割。
根据本发明的一些实施例，所述生成利于神经网络训练和测试的预处理图像包括：图像的预处理过程包括灰度变换处理、ROI提取处理、归一化处理、对比度受限的自适应直方图均衡处理和伽马调整处理。
根据本发明的一些实施例，所述生成利于神经网络训练的金标准包括：通过重复线追踪、宽线检测器、Gabor滤波器、最大曲率、平均曲率和增强型最大曲率来提取六种手指静脉分割的纹路图，并将六种所述纹路图附以不同的权重来得到神经网络训练的所述金标准。
根据本发明的一些实施例，所述将高斯成对势的密集条件随机场作为循环神经网络进行平均场近似推理包括：所述密集条件随机场的神经网络分解过程包括 初始化、消息传递、加权滤波器输出、兼容性变换、增加一元势和归一化步骤。
根据本发明的第二方面实施例的一种基于神经网络和概率图模型的手指静脉分割装置，包括：预处理图像生成单元，用于生成利于神经网络训练和测试的预处理图像；金标准生成单元，用于生成利于神经网络训练的金标准；平均场近似推理单元，用于将高斯成对势的密集条件随机场作为循环神经网络进行平均场近似推理；训练处理单元，用于将所述密集条件随机场融入至神经网络中，进行神经网络的训练处理。
根据本发明的第三方面实施例的一种基于神经网络和概率图模型的手指静脉分割设备，包括：至少一个控制处理器和用于与所述至少一个控制处理器通信连接的存储器；所述存储器存储有可被所述至少一个控制处理器执行的指令，所述指令被所述至少一个控制处理器执行，以使所述至少一个控制处理器能够执行如上述第一方面实施例的基于神经网络和概率图模型的手指静脉分割方法。
根据本发明的第四方面实施例的一种计算机可读存储介质，所述计算机可读存储介质存储有计算机可执行指令，所述计算机可执行指令用于使计算机执行如上述第一方面实施例的基于神经网络和概率图模型的手指静脉分割方法。
根据本发明的第五方面实施例的一种计算机程序产品，所述计算机程序产品包括存储在计算机可读存储介质上的计算机程序，所述计算机程序包括程序指令，当所述程序指令被计算机执行时，使计算机执行如上述第一方面实施例的基于神经网络和概率图模型的手指静脉分割方法。
本发明的附加方面和优点将在下面的描述中部分给出，部分将从下面的描述中变得明显，或通过本发明的实践了解到。
附图说明
本发明的上述和/或附加的方面和优点从结合下面附图对实施例的描述中将变得明显和容易理解，其中：
图1为本发明实施例的一种基于神经网络和概率图模型的手指静脉分割方法的步骤流程图；
图2为本发明实施例的一种基于神经网络和概率图模型的手指静脉分割方法的框架流程图；
图3为本发明实施例的一种基于神经网络和概率图模型的手指静脉分割方法中生成利于神经网络训练的金标准的流程图；
图4为本发明实施例的一种基于神经网络和概率图模型的手指静脉分割方法中密集条件随机场中的平均场分解为常见的CNN操作过程的示意图；
图5为本发明实施例的一种基于神经网络和概率图模型的手指静脉分割方法中将迭代平均场近似为循环神经网络示意图；
图6为本发明实施例的一种基于神经网络和概率图模型的手指静脉分割方法中密集CRF中的平均场分解为常见的CNN操作算法流程图；
图7为本发明实施例的一种基于神经网络和概率图模型的手指静脉分割装置的结构图；
图8为本发明实施例的一种基于神经网络和概率图模型的手指静脉分割设备的结构图。
附图标记：
基于神经网络和概率图模型的手指静脉分割装置100、预处理图像生成单元110、金标准生成单元120、平均场近似推理单元130、训练处理单元140；
基于神经网络和概率图模型的手指静脉分割设备200、控制处理器210、存储器220。
具体实施方式
下面详细描述本发明的实施例，所述实施例的示例在附图中示出，其中自始至终相同或类似的标号表示相同或类似的元件或具有相同或类似功能的元件。下面通过参考附图描述的实施例是示例性的，仅用于解释本发明，而不能理解为对本发明的限制。
在本发明的描述中，需要理解的是，涉及到方位描述，例如上、下、前、后、左、右等指示的方位或位置关系为基于附图所示的方位或位置关系，仅是为了便 于描述本发明和简化描述，而不是指示或暗示所指的装置或元件必须具有特定的方位、以特定的方位构造和操作，因此不能理解为对本发明的限制。
在本发明的描述中，若干的含义是一个或者多个，多个的含义是两个以上，大于、小于、超过等理解为不包括本数，以上、以下、以内等理解为包括本数。如果有描述到第一、第二只是用于区分技术特征为目的，而不能理解为指示或暗示相对重要性或者隐含指明所指示的技术特征的数量或者隐含指明所指示的技术特征的先后关系。
本发明的描述中，除非另有明确的限定，设置、安装、连接等词语应做广义理解，所属技术领域技术人员可以结合技术方案的具体内容合理确定上述词语在本发明中的具体含义。
参照图1，根据本发明的第一方面实施例的一种基于神经网络和概率图模型的手指静脉分割方法，包括以下步骤：
S1：生成利于神经网络训练和测试的预处理图像；
S2：生成利于神经网络训练的金标准；所述金标准即为标签；
S3：将高斯成对势的密集条件随机场作为循环神经网络进行平均场近似推理；
S4：将所述密集条件随机场融入至神经网络中，进行神经网络的训练处理。使用全卷积神经网络架构作为网络的第一部分，为条件随机场提供一元势能；将整幅图像(或多幅图像)作为小批次处理，利用适当的损失函数，计算网络中每个像素输出的误差。
根据本发明第一方面实施例的一种基于神经网络和概率图模型的手指静脉分割方法，至少具有如下有益效果：本发明的一种基于神经网络和概率图模型的手指静脉分割方法，能够将具有高斯成对势的密集条件随机场(CRF)作为一种循环神经网络(RNN)建立了平均场近似推理，它可以在前向传递过程中细化传统神经网络(CNN)的粗输出，同时在训练过程中将误差传递回神经网络(CNN)，经过一定时间的训练，由于充分利用了密集条件随机场(CRF)的优势，因此可 以产生更精细的静脉分割。
根据本发明的一些实施例，所述生成利于神经网络训练和测试的预处理图像包括：图像的预处理过程包括灰度变换处理、ROI提取处理、归一化处理、对比度受限的自适应直方图均衡处理和伽马调整处理。目的在于突出静脉纹路丰富的区域，减少处理时间，增加精度。
根据本发明的一些实施例，所述生成利于神经网络训练的金标准包括：通过重复线追踪、宽线检测器、Gabor滤波器、最大曲率、平均曲率和增强型最大曲率来提取六种手指静脉分割的纹路图，并将六种所述纹路图附以不同的权重来得到神经网络训练的所述金标准。该方法可以充分发挥每种传统提取纹路的优势，弥补劣势，使得标签尽量准确。
根据本发明的一些实施例，所述将高斯成对势的密集条件随机场作为循环神经网络进行平均场近似推理包括：所述密集条件随机场的神经网络分解过程包括初始化、消息传递、加权滤波器输出、兼容性变换、增加一元势和归一化步骤。
为了解决现有技术中无法达到良好的分割效果，传统的CNN具有感受野较大的卷积滤波器，重构后生成像素级标签会产生粗输出，且没有适于神经网络训练的参考标准等问题，本发明实施例以下结合附图对本发明进行进一步详细说明。
图2是本发明实施例的手指静脉分割方法总体框架流程图，包括图像预处理、平均场近似推理、标签制作和搭建网络四部分。
图像的预处理包括灰度变换、ROI提取、归一化、对比度受限的自适应直方图均衡和伽马调整等。由于单通道的静脉血管背景对比度优于RGB图像，故将原始RGB静脉图像转换为单通道图像，接着利用归一化和对比度受限的自适应直方图均衡来增强前后背景对比度，最后引入伽马调整，进一步提高图像质量。
图3是金标准的制作过程，在此选择了六种传统方法作为金标准(即标签)的基础构造分割图，根据分割质量的好坏以及各种方法的优劣附以不同的权重，权重的大小在图3中已标出，接着将其进行图像叠加、平均、二值化、滤波等操作，得到最终用于网络训练的金标准。
图4是密集条件随机场中的平均场分解为常见的CNN操作过程。图6是密集CRF中的平均场分解为常见的CNN操作算法流程图，包括初始化、消息传递、加权滤波器输出、兼容性变换、增加一元势和归一化步骤。下面详细介绍下该步骤。
初始化：
在算法初始化阶段，执行
Figure PCTCN2020078177-appb-000001
操作，其中Z i＝∑ lexp(U i(l))是归一化因子。该操作不包含任何参数，并且在执行softmax转换的通常向后传递计算后，可以将反向传播步骤中输出时接收到的误差微分传递到一元势输入。
消息传递：
Q(l)是一张概率图片，对应着原始图像每个像素点属于类别l的概率。k(f i，f j)是一个高斯核权重，这就相当于神经网络的卷积运算一样。具体k(f i，f j)一般包含：位置、颜色度量等作为特征。在密集CRF公式中，消息传递是通过对Q值应用M高斯滤波器来实现的。高斯滤波系数是根据像素位置和RGB值等图像特征得到的，这些特征反映了一个像素与其他像素之间的关系有多强。由于CRF可能是全连接的，每个滤波器的感受野横跨整个图像，因此不可能依靠蛮力实现滤波器。在反向传播过程中，通过将对滤波器输出的误差导数利用相同的M高斯滤波器反向发送，计算出对滤波器输入的误差导数。在Permutohedral lattice操作方面，这可以通过在模糊阶段只反转可分离滤波器的顺序来实现，同时以与前向通道相同的方式构建Permutohedral lattice，splatting和slicing。因此，通过这个滤波阶段的反向传播也可以在O(N)时间内进行。
加权滤波器输出：
这一步可以看成是采用1*1的卷积层，对多个特征图进行卷积层运算。经过运算完后，每两张特征图又输出一张新的概率图。平均场迭代的下一步是对前一步的M个滤波器输出(每个类标签l)取加权和。当每个类标签被单独考虑时，这可以被看作是与一个带有M个输入通道，一个输出通道的1×1滤波器的通常卷积。由于此步骤的输入和输出在反向传播期间都是已知的，因此可以计算出相对于滤波器权值的误差导数，从而使自动学习滤波器权值成为可能(前一阶段的每 个高斯滤波器输出的相对贡献)。对输入的误差导数也可以用通常的方法计算，从而将误差导数传递到前一阶段。为了获得更多的可调参数，我们对每个类标签使用独立的核权重。
兼容性变换：
在兼容性变换步骤中，根据这些标签之间的兼容性，在不同程度上共享前一步的输出(在算法1中由
Figure PCTCN2020078177-appb-000002
表示)。两个标签l和l′之间的兼容性由标签兼容性函数μ(l，l′)参数化。Potts模型，由μ(l，l′)＝[l≠l′]，其中[·]是艾佛森括号(方括号内的条件满足则为1，不满足则为0)，如果为具有相似属性的像素分配不同的标签，则分配固定的惩罚。该模型的局限性在于它为所有不同的标签对赋予相同的惩罚。直观地，通过考虑不同标签对之间的兼容性并相应地惩罚分配，可以获得更好的结果。因此，最好从数据中学习函数μ，而不是使用Potts模型预先修复函数。兼容性变换步骤可以看作是另一个卷积层，其中滤波器的空间感受野为1×1，输入和输出通道数均为L。学习该滤波器的权值相当于学习标签兼容性函数μ。由于这一步是一个常见的卷积运算，因此可以将误差微分从这一步的输出转移到输入。
增加一元势：
在这个步骤中，兼容性变换阶段的输出从一元输入U中减去对应元素。虽然此步骤中不涉及任何参数，但通过将此步骤输出处的差值复制到具有适当符号的两个输入，可以轻松地传输误差。
归一化：
最后，迭代的归一化步骤可以看作是另一种没有参数的softmax操作。此步骤输出处的微分可以使用softmax操作的反向传播传递到输入。
图5是将迭代平均场近似为循环神经网络示意图。下面具体介绍该实现流程。
使用函数f θ来表示通过一次平均场迭代完成的变换：给定图像I，像素对的一元势能值U和来自前一次迭代的边缘概率Q in的估计，以及在给定f θ(U，Q in，I) 的一次平均场迭代后边缘分布的下一个估计。向量Q＝{w (m)，μ(l，l′)}，m∈{1，...，M}，l，l′∈{l 1，...，l L}表示CRF参数(高斯核权重和兼容性函数)。这个近似过程如以下几个公式给出，其中T为平均场迭代次数。
Figure PCTCN2020078177-appb-000003
H 2(t)＝f θ(U，H 1(t)，I)，0≤t≤T      (2)
Figure PCTCN2020078177-appb-000004
在训练过程中，可以使用整幅图像(或多幅图像)作为小批次处理，利用适当的损失函数，如与图像的标签分割相关的softmax损失，计算网络中每个像素输出的误差。
参照图7，根据本发明的第二方面实施例的一种基于神经网络和概率图模型的手指静脉分割装置100，包括：
预处理图像生成单元110，用于生成利于神经网络训练和测试的预处理图像；
金标准生成单元120，用于生成利于神经网络训练的金标准；
平均场近似推理单元130，用于将高斯成对势的密集条件随机场作为循环神经网络进行平均场近似推理；
训练处理单元140，用于将所述密集条件随机场融入至神经网络中，进行神经网络的训练处理。
需要说明的是，由于本实施例中的基于神经网络和概率图模型的手指静脉分割装置100与上述的基于神经网络和概率图模型的手指静脉分割方法基于相同的发明构思，因此，方法实施例中的相应内容同样适用于本装置实施例，此处不再详述。
参照图8，根据本发明的第三方面实施例的一种基于神经网络和概率图模型的手指静脉分割设备200，该基于神经网络和概率图模型的手指静脉分割设备200可以是任意类型的智能终端，例如手机、平板电脑、个人计算机等。
具体地，该基于神经网络和概率图模型的手指静脉分割设备200包括：一个或多个控制处理器210和存储器220，图8中以一个控制处理器210为例。
控制处理器210和存储器220可以通过总线或者其他方式连接，图8中以通过总线连接为例。
存储器220作为一种非暂态计算机可读存储介质，可用于存储非暂态软件程序、非暂态性计算机可执行程序以及模块，如本发明实施例中的基于神经网络和概率图模型的手指静脉分割方法对应的程序指令/模块，例如，图7中所示的单元110-140。控制处理器210通过运行存储在存储器220中的非暂态软件程序、指令以及模块，从而执行基于神经网络和概率图模型的手指静脉分割装置100的各种功能应用以及数据处理，即实现上述方法实施例的基于神经网络和概率图模型的手指静脉分割方法。
存储器220可以包括存储程序区和存储数据区，其中，存储程序区可存储操作系统、至少一个功能所需要的应用程序；存储数据区可存储根据基于神经网络和概率图模型的手指静脉分割装置100的使用所创建的数据等。此外，存储器220可以包括高速随机存取存储器，还可以包括非暂态存储器，例如至少一个磁盘存储器件、闪存器件、或其他非暂态固态存储器件。在一些实施方式中，存储器220可选包括相对于控制处理器210远程设置的存储器220，这些远程存储器220可以通过网络连接至该基于神经网络和概率图模型的手指静脉分割设备200。上述网络的实例包括但不限于互联网、企业内部网、局域网、移动通信网及其组合。
所述一个或者多个模块存储在所述存储器220中，当被所述一个或者多个控制处理器210执行时，执行上述方法实施例中的基于神经网络和概率图模型的手指静脉分割方法，例如，执行以上描述的图1中的方法步骤S1至S4，实现图7中的单元110-140的功能。
根据本发明的第四方面实施例的一种计算机可读存储介质，所述计算机可读存储介质存储有计算机可执行指令，该计算机可执行指令被一个或多个控制处理 器210执行，例如，被图8中的一个控制处理器210执行，可使得上述一个或多个控制处理器210执行上述方法实施例中的基于神经网络和概率图模型的手指静脉分割方法，例如，执行以上描述的图1中的方法步骤S1至S4，实现图7中的单元110-140的功能。
以上所描述的装置实施例仅仅是示意性的，其中所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部模块来实现本实施例方案的目的。
通过以上的实施方式的描述，本领域技术人员可以清楚地了解到各实施方式可借助软件加通用硬件平台的方式来实现。本领域技术人员可以理解实现上述实施例方法中的全部或部分流程是可以通过计算机程序来指令相关的硬件来完成，所述的程序可存储于一计算机可读取存储介质中，该程序在执行时，可包括如上述方法的实施例的流程。其中，所述的存储介质可为磁碟、光盘、只读存储记忆体(ReadOnly Memory,ROM)或随机存储记忆体(Random Access Memory,RAM)等。
根据本发明的第五方面实施例的一种计算机程序产品，所述计算机程序产品包括存储在计算机可读存储介质上的计算机程序，所述计算机程序包括程序指令，当所述程序指令被计算机执行时，使计算机执行如上述第一方面实施例的基于神经网络和概率图模型的手指静脉分割方法。
上面结合附图对本发明实施例作了详细说明，但是本发明不限于上述实施例，在所述技术领域普通技术人员所具备的知识范围内，还可以在不脱离本发明宗旨的前提下作出各种变化。

Claims (7)

  1. 一种基于神经网络和概率图模型的手指静脉分割方法，其特征在于，包括以下步骤：
    生成利于神经网络训练和测试的预处理图像；
    生成利于神经网络训练的金标准；
    将高斯成对势的密集条件随机场作为循环神经网络进行平均场近似推理；
    将所述密集条件随机场融入至神经网络中，进行神经网络的训练处理。
  2. 根据权利要求1所述的基于神经网络和概率图模型的手指静脉分割方法，其特征在于，所述生成利于神经网络训练和测试的预处理图像包括：图像的预处理过程包括灰度变换处理、ROI提取处理、归一化处理、对比度受限的自适应直方图均衡处理和伽马调整处理。
  3. 根据权利要求1所述的基于神经网络和概率图模型的手指静脉分割方法，其特征在于，所述生成利于神经网络训练的金标准包括：通过重复线追踪、宽线检测器、Gabor滤波器、最大曲率、平均曲率和增强型最大曲率来提取六种手指静脉分割的纹路图，并将六种所述纹路图附以不同的权重来得到神经网络训练的所述金标准。
  4. 根据权利要求1所述的基于神经网络和概率图模型的手指静脉分割方法，其特征在于，所述将高斯成对势的密集条件随机场作为循环神经网络进行平均场近似推理包括：所述密集条件随机场的神经网络分解过程包括初始化、消息传递、加权滤波器输出、兼容性变换、增加一元势和归一化步骤。
  5. 一种基于神经网络和概率图模型的手指静脉分割装置，其特征在于，包括：
    预处理图像生成单元，用于生成利于神经网络训练和测试的预处理图像；
    金标准生成单元，用于生成利于神经网络训练的金标准；
    平均场近似推理单元，用于将高斯成对势的密集条件随机场作为循环神经网络进行平均场近似推理；
    训练处理单元，用于将所述密集条件随机场融入至神经网络中，进行神经网络的训练处理。
  6. 一种基于神经网络和概率图模型的手指静脉分割设备，其特征在于，包括：至少一个控制处理器和用于与所述至少一个控制处理器通信连接的存储器；所述存储器存储有可被所述至少一个控制处理器执行的指令，所述指令被所述至少一个控制处理器执行，以使所述至少一个控制处理器能够执行如权利要求1至4任一项所述的基于神经网络和概率图模型的手指静脉分割方法。
  7. 一种计算机可读存储介质，其特征在于：所述计算机可读存储介质存储有计算机可执行指令，所述计算机可执行指令用于使计算机执行如权利要求1至4任一项所述的基于神经网络和概率图模型的手指静脉分割方法。
PCT/CN2020/078177 2019-09-27 2020-03-06 基于神经网络和概率图模型的手指静脉分割方法及装置 Ceased WO2021056972A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910921732.7 2019-09-27
CN201910921732.7A CN110826563B (zh) 2019-09-27 2019-09-27 基于神经网络和概率图模型的手指静脉分割方法及装置

Publications (1)

Publication Number Publication Date
WO2021056972A1 true WO2021056972A1 (zh) 2021-04-01

Family

ID=69548636

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/078177 Ceased WO2021056972A1 (zh) 2019-09-27 2020-03-06 基于神经网络和概率图模型的手指静脉分割方法及装置

Country Status (2)

Country Link
CN (1) CN110826563B (zh)
WO (1) WO2021056972A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113505716A (zh) * 2021-07-16 2021-10-15 重庆工商大学 静脉识别模型的训练方法、静脉图像的识别方法及装置
CN113870214A (zh) * 2021-09-26 2021-12-31 盛视科技股份有限公司 静脉图像的质量评估方法、评估装置及终端
CN115965551A (zh) * 2022-12-30 2023-04-14 上海联影医疗科技股份有限公司 医学影像处理方法、系统及存储介质
CN119850583A (zh) * 2025-01-06 2025-04-18 南京大学 一种人体背部棘突线提取方法、设备、介质及产品

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110826563B (zh) * 2019-09-27 2023-11-24 五邑大学 基于神经网络和概率图模型的手指静脉分割方法及装置
CN115222681A (zh) * 2022-06-29 2022-10-21 电子科技大学 一种用于分割医学图像中的对象的方法和数据处理单元

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180068429A1 (en) * 2015-04-15 2018-03-08 Institute Of Automation Chinese Academy Of Sciences Image Steganalysis Based on Deep Learning
CN110176006A (zh) * 2019-05-15 2019-08-27 北京航空航天大学 图像前景物体分割方法及装置
CN110826563A (zh) * 2019-09-27 2020-02-21 五邑大学 基于神经网络和概率图模型的手指静脉分割方法及装置

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107256550A (zh) * 2017-06-06 2017-10-17 电子科技大学 一种基于高效cnn‑crf网络的视网膜图像分割方法
CN108876796A (zh) * 2018-06-08 2018-11-23 长安大学 一种基于全卷积神经网络和条件随机场的道路分割系统及方法
CN109671076A (zh) * 2018-12-20 2019-04-23 上海联影智能医疗科技有限公司 血管分割方法、装置、电子设备及存储介质

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180068429A1 (en) * 2015-04-15 2018-03-08 Institute Of Automation Chinese Academy Of Sciences Image Steganalysis Based on Deep Learning
CN110176006A (zh) * 2019-05-15 2019-08-27 北京航空航天大学 图像前景物体分割方法及装置
CN110826563A (zh) * 2019-09-27 2020-02-21 五邑大学 基于神经网络和概率图模型的手指静脉分割方法及装置

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
ZHENG SHUAI; JAYASUMANA SADEEP; ROMERA-PAREDES BERNARDINO; VINEET VIBHAV; SU ZHIZHONG; DU DALONG; HUANG CHANG; TORR PHILIP H. S.: "Conditional Random Fields as Recurrent Neural Networks", 2015 IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION (ICCV), IEEE, 7 December 2015 (2015-12-07), pages 1529 - 1537, XP032866501, DOI: 10.1109/ICCV.2015.179 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113505716A (zh) * 2021-07-16 2021-10-15 重庆工商大学 静脉识别模型的训练方法、静脉图像的识别方法及装置
CN113505716B (zh) * 2021-07-16 2022-07-01 重庆工商大学 静脉识别模型的训练方法、静脉图像的识别方法及装置
CN113870214A (zh) * 2021-09-26 2021-12-31 盛视科技股份有限公司 静脉图像的质量评估方法、评估装置及终端
CN115965551A (zh) * 2022-12-30 2023-04-14 上海联影医疗科技股份有限公司 医学影像处理方法、系统及存储介质
CN119850583A (zh) * 2025-01-06 2025-04-18 南京大学 一种人体背部棘突线提取方法、设备、介质及产品
CN119850583B (zh) * 2025-01-06 2025-09-26 南京大学 一种人体背部棘突线提取方法、设备、介质及产品

Also Published As

Publication number Publication date
CN110826563B (zh) 2023-11-24
CN110826563A (zh) 2020-02-21

Similar Documents

Publication Publication Date Title
CN110826563B (zh) 基于神经网络和概率图模型的手指静脉分割方法及装置
CN113705769B (zh) 一种神经网络训练方法以及装置
Shan et al. Automatic facial expression recognition based on a deep convolutional-neural-network structure
CN109815826B (zh) 人脸属性模型的生成方法及装置
CN110070107B (zh) 物体识别方法及装置
US20240404018A1 (en) Image processing method and apparatus, device, storage medium and program product
CN109359608B (zh) 一种基于深度学习模型的人脸识别方法
CN109949255A (zh) 图像重建方法及设备
CN106127120A (zh) 姿势估计方法和装置、计算机系统
CN113158943A (zh) 一种跨域红外目标检测方法
CN112861718A (zh) 一种轻量级特征融合人群计数方法及系统
CN115862119B (zh) 基于注意力机制的人脸年龄估计方法及装置
CN113903063A (zh) 基于深度时空网络决策融合的人脸表情识别方法及系统
CN114240955A (zh) 一种半监督的跨领域自适应的图像分割方法
CN116802683A (zh) 图像的处理方法和系统
CN110135435B (zh) 一种基于广度学习系统的显著性检测方法及装置
Yang et al. A Face Detection Method Based on Skin Color Model and Improved AdaBoost Algorithm.
Wang et al. MetaScleraSeg: an effective meta-learning framework for generalized sclera segmentation
Yang et al. Retinal vessel segmentation based on an improved deep forest
CN110751057A (zh) 基于长短时记忆循环神经网络的手指静脉验证方法及装置
Zhang et al. Accuracy and robustness evaluation of deep learning algorithms in facial recognition systems
Liang et al. Recognition and classification of ornamental fish image based on machine vision
CN117523626A (zh) 伪rgb-d人脸识别法
CN120635457A (zh) 一种双分割头频率解耦学习与熵变化伪标签筛选相结合的半监督分割方法
CN112016592A (zh) 基于交叉领域类别感知的领域适应语义分割方法及装置

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20869871

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20869871

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 20869871

Country of ref document: EP

Kind code of ref document: A1