WO2017152403A1 - 一种基于mcmc框架下的子超图匹配方法和装置 - Google Patents

一种基于mcmc框架下的子超图匹配方法和装置 Download PDF

Info

Publication number
WO2017152403A1
WO2017152403A1 PCT/CN2016/076030 CN2016076030W WO2017152403A1 WO 2017152403 A1 WO2017152403 A1 WO 2017152403A1 CN 2016076030 W CN2016076030 W CN 2016076030W WO 2017152403 A1 WO2017152403 A1 WO 2017152403A1
Authority
WO
WIPO (PCT)
Prior art keywords
sub
matching
supergraph
hypergraph
state
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/CN2016/076030
Other languages
English (en)
French (fr)
Inventor
王文敏
张若楠
王荣刚
李革
董胜富
王振宇
李英
赵辉
高文
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.)
Peking University Shenzhen Graduate School
Original Assignee
Peking University Shenzhen Graduate School
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 Peking University Shenzhen Graduate School filed Critical Peking University Shenzhen Graduate School
Priority to PCT/CN2016/076030 priority Critical patent/WO2017152403A1/zh
Priority to US16/079,660 priority patent/US11347979B2/en
Publication of WO2017152403A1 publication Critical patent/WO2017152403A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/29—Graphical models, e.g. Bayesian networks
    • G06F18/295—Markov models or related models, e.g. semi-Markov models; Markov random fields; Networks embedding Markov models
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/01—Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N7/00—Computing arrangements based on specific mathematical models
    • G06N7/01—Probabilistic graphical models, e.g. probabilistic networks

Definitions

  • the present application relates to the field of computer vision and image recognition technologies, and in particular, to a sub-supergraph matching method and apparatus based on MCMC (Markov Chain Monte Carlo) framework.
  • MCMC Markov Chain Monte Carlo
  • feature matching has a wide range of applications, such as target detection, target recognition, tracking, and so on. Because in the real picture, there are often some complicated changes, such as background clutter, deformation, repeated patterns, etc., which will make the information content of the feature descriptors insufficient.
  • Hypergraph matching provides a powerful representation model for feature matching by describing the relationship between feature points and features. It not only saves the same nodes and edges as the graph, but most importantly it increases.
  • the reference quantity of the geometric position information of the object is an extension of the high-dimensional space. Therefore, the feature matching by the super-picture structure is particularly prominent.
  • Hypergraph matching is widely used in network topology, big data analysis, target recognition and tracking, 3D reconstruction, scene understanding, video analysis, etc., especially for non-rigid objects and image matching for deformation. Therefore, its research is very important.
  • IQP integer programming problem
  • the super-picture matching method in the prior art still has many problems when performing image feature matching, and needs to be improved.
  • the present application provides a sub-supergraph matching method and device based on the MCMC framework, which solves the problems of low feature matching efficiency and insufficient accuracy in the prior art.
  • the present application provides a sub-super map matching method based on the MCMC framework, including:
  • the Markov chain in the super-map matching is constructed, and when the Markov chain reaches equilibrium, the sub-supergraph matching result is output according to the objective function.
  • the present application provides a sub-supergraph matching device based on the MCMC framework, including:
  • a sub-supergraph building module configured to construct a sub-supergraph, the first sub-supergraph located in the original image and the second sub-supergraph located in the image to be matched;
  • a target distribution function building module and constructing a target distribution function based on the MCMC framework
  • the matching result output module is configured to construct a Markov chain in the sub-supergraph matching, and when the Markov chain reaches equilibrium, the sub-supergraph matching result is output according to the objective function.
  • the present invention provides a sub-supergraph matching method and device based on the MCMC framework, and proposes matching object features by constructing a sub-supergraph.
  • the sub-supergraph is used to express the appearance of the object and Position information for object feature matching enables faster matching of images with higher speed and accuracy.
  • sub-supergraphs have advantages over graphs and supergraphs.
  • the sub-supergraph has more geometric information than the graph (for example, angle transformation, rotation, scale, etc.), has lower difficulty than the hypergraph, and has better ductility.
  • the method and device have strong anti-interference ability, good robustness, and can adapt to more complicated scenarios, especially for outliers.
  • FIG. 1 is a schematic flowchart of a sub-supergraph matching method based on an MCMC framework in an embodiment of the present application
  • FIG. 2 is a schematic flowchart of a sub-super map matching method based on an MCMC framework in an embodiment of the present application
  • FIG. 3 is a schematic diagram of a process of adding and deleting states in a state transition based on a sub-supergraph matching method in an MCMC framework according to an embodiment of the present application;
  • FIG. 4-9 is a schematic diagram showing a comparison between a recall rate, an accuracy rate, and an F value obtained by performing image matching using the method provided in this embodiment and other methods in the prior art;
  • FIG. 4-9 is a schematic diagram showing a comparison between a recall rate, an accuracy rate, and an F value obtained by performing image matching using the method provided in this embodiment and other methods in the prior art;
  • FIG. 10 is a schematic diagram of a module of a sub-hypergraph matching device based on an MCMC framework according to an embodiment of the present application.
  • the feature matching problem can be regarded as a hypergraph matching problem.
  • the hyper graph can not only represent the information between the points in the image, but also include the occurrence of various edges (such as a hyper-edge set of 3 points, 4 A super-edge set consisting of points, etc.).
  • the sub-supergraph is a subset of the hypergraph, that is, the hypergraph with fewer nodes than the hypergraph.
  • the subsupergraph matching is to find the best matching structure common to the two graphs.
  • the sub-hypergraph has the same valuable geometric information as the hypergraph, and at the same time has a lower complexity than the hypergraph, and the sub-supergraph matching is not the same as the matching point in the pair of images that must be required in the hypergraph matching. For example, two supergraphs, A and B, 20 points in A match 20 points in B, then one to one match is matched. In the super-graph, it may only correspond to the matching pairs of the 18 most similar points, instead of 20 one-to-one correspondence.
  • this embodiment provides a sub-super map matching method based on the MCMC framework, including the following steps:
  • Step 1.1 Construct a sub-supergraph.
  • the sub-supergraph includes a first sub-supergraph located in the original image and a second sub-supergraph located in the image to be matched.
  • Step 1.2 Build the objective function.
  • Step 1.3 Construct a target distribution function based on the MCMC framework.
  • Step 1.4 Constructing a Markov chain in the super-map matching, and outputting the super-map matching result according to the objective function when the Markov chain reaches equilibrium.
  • Step 2.1 Construct a sub-supergraph. (first sub-picture)
  • G P (V P , E P , A P )
  • G Q (V Q , E Q , A Q )
  • G P contains n P points
  • G Q contains n Q points.
  • V represents the set of points
  • E represents the set of super-edges
  • A represents the attribute information of the super-edge.
  • a super edge allows the "edge" of any number of vertices to be connected, ie the super edge is an edge with n points joined together.
  • n takes 3, that is, V is a set of edges in which all 3 points are connected. In other embodiments, the value of n can be set according to actual needs.
  • Step 2.2 Build a generic objective function.
  • the similarity function is used. To show that it is used to measure the super-edge in the sub-supergraph of two d dimensions versus The similarity between the two. Therefore, the overall similarity of the d-dimensional hypergraph is expressed by the following formula (2), and ⁇ d is the weight of the d-dimensional spatial similarity H.
  • the hypergraph matching score can be described as the sum of the similarities of the d-dimensional superedges, as shown in the following formula (3):
  • the similarity measurement method here is different from the conventional method.
  • the dimension of the sub-supergraph more than 2 dimensions is considered to be a sub-supergraph.
  • the dimension d 3.
  • the similarity tensor of O.Duchenne The physical meaning is to find a graph based on the point information in the hypergraph (feature points in the image). All the triangles in the middle, such as information such as points, edges, and angular rotations, are matched as tensors such as triangles.
  • this embodiment adopts the following method:
  • the number of points in the first sub-supergraph is greater than the first threshold, determining that the number of triangles is the product of the number of points in the two sub-supergraphs, ie, n P *n Q ; otherwise, determining the number of triangles as the first sub-super
  • the maximum number of triangles that can be formed by all points in the graph that is, the number of combinations of 3 points in any of the n P points of G P Let's use this as a tuple to find the set of points in G Q that correspond to it.
  • the first threshold is designed according to the actual situation.
  • the number of triangles is simply determined as the product of the number of points in the two sub-supergraphs, a combination of n P *n Q triangles will be randomly selected, and a very important set of triangle combination points may be missed; the number of triangles is determined as The maximum number of triangles that can be formed by all points in the first sub-supergraph is considered more comprehensive and more comprehensive. However, when the number of points is larger than a certain value, the consideration of thinning is very important, otherwise a lot of time is wasted, so in this case the number of triangles can be determined as the product of the number of points in the two sub-supergraphs.
  • the number of triangles is determined to be the maximum number of triangles that can be formed by all the points in the first sub-supergraph, if the conditions are allowed by the device or if there is a large requirement for the speed, it is the best choice. .
  • the method of constructing a triangle adopted in this embodiment can improve the matching efficiency under the premise of ensuring the matching accuracy.
  • Step 2.3 Obtain the prior knowledge parameters by learning, and add the prior knowledge parameters to the ordinary objective function.
  • This embodiment proposes to construct a sub-supergraph matching method based on the meta-algorithm under the framework of Markov Monte Carlo (MCMC).
  • this embodiment preferably integrates the a priori strategy into the original.
  • the problem (formula (4)).
  • This embodiment adds a reliable prior knowledge to the original problem (formula (4)), that is, the second term ⁇ T D(
  • ⁇ is obtained through the learning method of graphs, and its physical meaning is the degree of compactness, that is, the probability density, which is also a kind of weight, which is determined by the similarity measure method and the size of the sub-supergraph; D(
  • Step 2.4 Construct a target distribution function based on the MCMC framework.
  • the state space defined in the MCMC framework is Y, and each state y ⁇ Y has a target distribution and a probability density of ⁇ (y). Because the probability density is proportional to the target distribution, one of them can be represented by one. , as shown in the following formula (6). Among them, the annealing simulation value T is used to control the variation of the distribution. f(y) is the value that makes the hypergraph match score Score(x) the largest in the current y state.
  • Score(x) is the hypergraph matching score
  • ) is the learned prior knowledge parameter
  • ⁇ is the probability density
  • f(y) is the value that maximizes the hypergraph match score Score(x) in the current y state
  • T is the annealing simulation value.
  • Step 2.5 Initialize each parameter, for example to initialize the annealing simulation value T.
  • the Markov chain in the super-map matching is constructed, that is, step 1.4 in FIG. 1, which specifically includes the following sub-steps.
  • Step 2.6 Determine whether the Markov chain reaches equilibrium, and whether the annealing simulation value T is greater than a preset threshold Tf, and if yes, go to step 2.7, otherwise go to step 2.8.
  • Step 2.7 When the Markov chain reaches equilibrium and the annealing simulation value T is greater than the preset threshold Tf, it indicates that the best matching pair obtained has met the requirements, and therefore, the matching result is output.
  • Step 2.8 Perform mode selection (random, increase, delete) and calculate the probability distribution in this mode.
  • this embodiment uses the Metropolis-Hastings algorithm.
  • the state transition kernel function is shown in the following equation (7), where ⁇ (mode) is the probability density of mode. Mode has three modes: random, add, and delete. For different modes, there is a different probability distribution q(y, y') from the current state y to the next state y'.
  • ⁇ (add) is the probability density of the added mode
  • ⁇ (delete) is the probability density of the deleted mode.
  • the probability distributions in the three modes of random, add, and delete are (other cases have a probability distribution of 0):
  • Step 2.9 Calculate the acceptance rate ⁇ from the current state y to the next state y', which is calculated as shown in equation (9):
  • Step 2.10 Comparing the calculated acceptance rate ⁇ with a preset threshold to determine whether the acceptance rate ⁇ reaches a threshold value to determine whether to accept the next state y' matching information and the hypergraph matching score, and if so, go to Step 2.11, otherwise go to step 2.13.
  • Step 2.11 When the acceptance rate ⁇ reaches the preset threshold, it is further determined whether the hypergraph matching score is improved after the state transition. If the state transition is accepted, go to step 2.12, otherwise go to step 2.13.
  • Step 2.12 Accept the state transition, update the best match pair and the hypergraph match score, that is, the best match pair information and the hypergraph match score of the sub-supergraph are all updated to the best match pair information and hypergraph in the y' state. Match score.
  • Step 2.13 Regardless of whether or not the state transition is accepted, in this step, it is determined whether the number of samples is greater than a predetermined threshold, and if yes, go to step 2.14, otherwise go to step 2.6.
  • Step 2.14 Update the annealing simulation value T.
  • the annealing simulation value is T
  • the number of samples is N
  • the initial values of the two are 0.
  • the threshold corresponding to the number of samples is Nmax (ie, the sample threshold)
  • the threshold corresponding to the annealing analog value is Tf (ie, the end value of the annealing simulation).
  • Tf the end value of the annealing simulation.
  • y' is sampled every time, and the number of samples N is increased by one.
  • sub-supergraph matching method based on the MCMC framework provided by this embodiment, it is proposed to perform object feature matching by constructing a sub-supergraph.
  • the sub-supergraph is used to express the appearance of the object and Position information for object feature matching enables faster matching of images with higher speed and accuracy.
  • sub-supergraphs have advantages over graphs and supergraphs.
  • the sub-supergraph has more geometric information than the graph (for example, angle transformation, rotation, scale, etc.), has lower difficulty than the hypergraph, and has better ductility.
  • the method and device have strong anti-interference ability, good robustness, and can adapt to more complicated scenarios, especially for outliers.
  • the matching between the points is performed one by one, instead of all the matching is completed, and the discretization operation is used to truly implement one-to-one matching, thereby making the matching more accurate and closer to matching.
  • a higher F means that the matching result is better.
  • FIG. 4-9 is a schematic diagram of comparison between the recall rate, the accuracy rate, and the F value obtained when the image provided by the embodiment is matched with other methods in the prior art.
  • the embodiment provides a sub-supergraph matching device based on the MCMC framework, including a sub-super graph building module 101, an objective function building module 102, and a target distribution function.
  • the building module 103 and the matching result output module 104 are constructed.
  • the sub-hypergraph construction module 101 is configured to construct a sub-supergraph including a first sub-supergraph located in the original image and a second sub-supergraph located in the image to be matched.
  • the objective function building block 102 is used to construct an objective function.
  • the target distribution function construction module 103 constructs a target distribution function based on the MCMC framework.
  • the matching result output module 104 is configured to construct a Markov chain in the sub-supergraph matching, and when the Markov chain reaches equilibrium, the sub-supergraph matching result is output according to the objective function.
  • the matching result output module 104 is configured to use the triangle tensor as a tuple for matching in the sub-supergraph matching process, and construct a triangle tensor in the following manner: when the number of points in the first sub-supergraph is greater than At the first threshold, it is determined that the number of triangles is the product of the number of points in the two sub-supergraphs; otherwise, the number of triangles is determined to be the maximum number of triangles that all points in the first sub-supergraph can constitute.
  • the objective function construction module 102 is configured to obtain a priori knowledge parameters by learning, and add prior knowledge parameters to the objective function.
  • the target distribution function construction module 103 is configured to define a state space Y in the MCMC framework, each state y ⁇ Y, a target distribution of ⁇ (y), and:
  • Score(x) is the hypergraph matching score
  • ) is the learned prior knowledge parameter
  • ⁇ is the probability density
  • ) Is the Euclidean distance of f(y)
  • f(y) is the maximum value of the supergraph matching score Score(x) in the current y state
  • T is the annealing simulation value
  • the state transition kernel function is:
  • ⁇ (mode) is the probability density of mode, and mode includes random, add, delete Three modes, q(y, y') are probability distributions from the current state y to the next state y';
  • the device provided in this embodiment corresponds to the method provided in the first embodiment, and details of the specific execution steps of the device are not described herein.
  • the program may be stored in a computer readable storage medium, and the storage medium may include: a read only memory. Random access memory, disk or optical disk, etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Mathematical Analysis (AREA)
  • Computational Mathematics (AREA)
  • Pure & Applied Mathematics (AREA)
  • Mathematical Optimization (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Algebra (AREA)
  • Evolutionary Biology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Medical Informatics (AREA)
  • Computational Linguistics (AREA)
  • Databases & Information Systems (AREA)
  • Operations Research (AREA)
  • Image Analysis (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Multimedia (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种基于MCMC框架下的子超图匹配方法和装置，提出通过构建子超图的方式进行物体特征的匹配。由于在大量的真实图片、视频中，物体都是不断变换，包含各种噪点，以及其他一些干扰因素的，以至于图像物体匹配检索变得非常困难，而通过子超图来表现物体的外观以及位置信息来进行物体特征匹配，就能更高速、更准确地实现图像的正确匹配。并且，子超图比图、超图更有优势。一方面，子超图比图具有更多的几何信息(例如，角度变换，旋转，尺度等)，比超图的具有更低的难度，具有更好的延展性。另一方面，该方法和装置具有较强的抗干扰能力，鲁棒性好，能适应较复杂的场景，尤其是针对异常值。

Description

一种基于MCMC框架下的子超图匹配方法和装置 技术领域
本申请涉及计算机视觉与图像识别技术领域，尤其涉及一种基于MCMC(Markov Chain Monte Carlo，马尔科夫蒙特卡洛)框架下的子超图匹配方法和装置。
背景技术
在计算机视觉研究中，特征匹配有着非常广泛的应用，比如说目标检测、目标识别、跟踪等等。由于在真实的图片中，经常会有一些复杂的变化，像背景杂波、形变、出现重复的模式等，都会使得特征描述子的信息量变得不充足。而超图匹配通过对特征点和特征间的关系的描述，为特征匹配提供了一个强有力的表现模型，其不仅可以保存和图一样的结点和边的信息，最重要的是其增加了物体几何位置信息的参考量，更是对高维空间的一个拓展，因此用超图结构来做特征匹配就显得尤为突出。
超图匹配广泛应用于网络拓扑图、大数据分析、目标识别和跟踪、三维重建、场景理解、视频分析等领域，尤其对非刚性物体以及发生形变的图像匹配有着更加广泛的通用性。因此对其的研究显得非常重要。但是由于其在数学上是二次分配问题，也就是NP难(non-deterministic polynomial，非确定性多项式)问题，故有很多接近最优解决问题的方案。而且大部分的方案都着重于提高匹配的召回率，极少的考虑匹配的准确率，这就导致在匹配的过程中带入了更多的噪点信息。同时，将其应用于特征匹配问题中时，整数规划问题(IQP)都会带来下面两个基本问题：
(1)在大多数的解决方案中都会预先确定好图中异常值(outliers)的具体个数，但其并没有考虑到解决方案中可能出现更多的异常值。以至于仅仅考虑召回率(recall)来衡量精确度(accuracy)是不够的，因为准确率(precision)并没有被考虑在内。
(2)同时在大多数的解决方案中，存在的异常值使得目标函数缺失了更精确的区分性。因为在大量实验中发现，当异常值出现时，目标函数不仅降低了准确度，而且对于召回率也是更加倾向于选择包含更多异常值的解决方案。因此，为了得到最近目标函数值，所需要的不仅是识别出正确的点(inliers)，而且需要排除出现的异常值(outliers)。
所以，现有技术中的超图匹配方法在进行图像的特征匹配时，还依旧存在着诸多的问题，需要进行改进。
发明内容
本申请提供了一种基于MCMC框架下的子超图匹配方法和装置，解决了现有技术中特征匹配效率低、精确度不够等问题。
根据本申请的第一方面，本申请提供了一种基于MCMC框架下的子超图匹配方法，包括：
构建子超图，所述子超图包括位于原始图像中的第一子超图和位于待匹配图像中的第二子超图；
构建目标函数；
构建基于MCMC框架下的目标分布函数；
构建子超图匹配中的马尔科夫链，在马尔科夫链达到平衡时，根据所述目标函数输出子超图匹配结果。
根据本申请的第二方面，本申请提供了一种基于MCMC框架下的子超图匹配装置，包括：
子超图构建模块，用于构建子超图，所述子超图包括位于原始图像中的第一子超图和位于待匹配图像中的第二子超图；
目标函数构建模块，用于构建目标函数；
目标分布函数构建模块，构建基于MCMC框架下的目标分布函数；
匹配结果输出模块，用于构建子超图匹配中的马尔科夫链，在马尔科夫链达到平衡时，根据所述目标函数输出子超图匹配结果。
本申请提供的一种基于MCMC框架下的子超图匹配方法和装置，提出通过构建子超图的方式进行物体特征的匹配。由于在大量的真实图片、视频中，物体都是不断变换，包含各种噪点，以及其他一些干扰因素的，以至于图像物体匹配检索变得非常困难，而通过子超图来表现物体的外观以及位置信息来进行物体特征匹配，就能更高速、更准确地实现图像的正确匹配。并且，子超图比图、超图更有优势。一方面，子超图比图具有更多的几何信息(例如，角度变换，旋转，尺度等)，比超图的具有更低的难度，具有更好的延展性。另一方面，该方法和装置具有较强的抗干扰能力，鲁棒性好，能适应较复杂的场景，尤其是针对异常值。
附图说明
图1为本申请一种实施例中基于MCMC框架下的子超图匹配方法的流程示意图；
图2为本申请一种实施例中基于MCMC框架下的子超图匹配方法的具体流程示意图；
图3为本申请一种实施例基于MCMC框架下的子超图匹配方法中状态转换中的添加和删除过程示意图；
图4-9为采用本实施例提供的方法与现有技术中的其他方法进行图像匹配时得到的召回率、准确率以及F值之间的比较示意图；
图10为本申请一种实施例中基于MCMC框架下的子超图匹配装置的模块示意图。
具体实施方式
本申请的发明构思在于：特征匹配问题可以看作是超图匹配问题。与图相比，超图(hyper graph)不仅仅能够表示图像中的点之间的信息，也能够包含各种边出现的情况(比如3个点组成的超边(hyper-edge)集合，4个点组成的超边集合等)。而子超图就是超图的一个子集，也就是节点比超图要少的超图，子超图匹配就是寻找两个图中共同有的最好的匹配结构。子超图具有和超图一样的有价值的几何信息，同时又比超图复杂度低，并且子超图匹配不像超图匹配中必须要求在一对图像中的匹配点是一样的。例如A和B两个超图，A中20个点匹配B中20个点，那么要一一对应匹配起来。子超图中，可能只对应其中最相似的18个点的匹配对，而不是20个都要一一对应。
下面通过具体实施方式结合附图对本申请作进一步详细说明。
实施例一
请参考图1，本实施例提供了一种基于MCMC框架下的子超图匹配方法，包括下面步骤：
步骤1.1：构建子超图。其中，子超图包括位于原始图像中的第一子超图和位于待匹配图像中的第二子超图。
步骤1.2：构建目标函数。
步骤1.3：构建基于MCMC框架下的目标分布函数。
步骤1.4：构建子超图匹配中的马尔科夫链，在马尔科夫链达到平衡时，根据该目标函数输出子超图匹配结果。
请参考图2，下面对步骤1.1-步骤1.4做进一步说明。
本实施例提供的基于MCMC框架下的子超图匹配方法的具体步骤如下。
步骤2.1：构建子超图。(第一子超图)
给定两个子超图GP(第一子超图)和GQ(第二子超图)，并且，GP＝(VP,EP,AP)，GQ＝(VQ,EQ,AQ)，GP含有nP个点，GQ含有nQ个点。其中V代表点的集合，E代表超边的集合，A代表超边的属性信息。一条超边允许连接任意多个顶点的“边”，即超边是有n个点连成的边。本实施例中，n取3，也就是V为所有3个点连成的边的集合。在其他实施例中，n的取值可以根据实际需求设定。
用X作为两个子超图的邻接矩阵。如果Xpq等于1，代表在VP中的第p个点与VQ中的第q个点是匹配的；否则Xpq等于0。描述成常规，如下面式(1)所示：
Figure PCTCN2016076030-appb-000001
步骤2.2：构建普通目标函数。
在子超图匹配中，相似度函数用
Figure PCTCN2016076030-appb-000002
来表示，它是用来衡量两个d维度的子超图中的超边
Figure PCTCN2016076030-appb-000003
与
Figure PCTCN2016076030-appb-000004
之间的相似度的。因此d维超图的整体的相似度用下面的公式(2)来表示，γd为d维空间相似度H的权重。
Figure PCTCN2016076030-appb-000005
Figure PCTCN2016076030-appb-000006
因此超图匹配评分就可以描述成d维超边的相似度总和，如下面的公式(3)所示：
Figure PCTCN2016076030-appb-000007
因此，需要找到最好的匹配对，也就是使得上面公式(3)中超图匹配得分最高的匹配对x*，如下面的公式(4)所示。
Figure PCTCN2016076030-appb-000008
但是这个问题是NP难的，也是整数二次方程编程问题(IQP)无法找到最优解的，因此如何找到一个适合的解决方案就非常重要。
本实施例中，对于此处的相似度衡量方式与传统方法不同。对于子超图的维度，大于2维就认为是子超图，为了简便起见，令维度d＝3。对于构建超图中超边(hyper-edge)的相似度衡量，可以使用O.Duchenne的相似度张量来衡量，其物理意义在于根据超图中的点信息(图像中为特征点)找到一个图中所有的三角形，这样包含点、边以及角度旋转等信息，以三角形这样的张量作为元组去进行匹配。
在构建三角形的过程中，本实施例采用下面方式：
当第一子超图中点的数目大于第一阈值时，确定三角形的数量为两个子超图中点的数目的乘积，即nP*nQ；否则，确定三角形的数量为第一子超图中所有点能够构成的三角形的数量的最大值，也就是在GP的nP个点中任意找出3个点的组合数
Figure PCTCN2016076030-appb-000009
来以此为元组去找到GQ中与之 对应的点集合。第一阈值根据实际情况设计。
如果单纯地将三角形的数量确定为两个子超图中点的数目的乘积，会随机选取nP*nQ个三角形的组合，可能会遗漏非常重要的三角形组合点集；将三角形的数量确定为第一子超图中所有点能够构成的三角形的数量的最大值，则考虑了所有情况，更加全面。然而，当点数大于一定值的时候，稀疏化的考虑是非常重要的，否则会浪费大量的时间，因此在这种情况下可以将三角形的数量确定为两个子超图中点的数目的乘积。当然，如果在设备允许的条件下，或不对速度有较大要求的情况下，将三角形的数量确定为第一子超图中所有点能够构成的三角形的数量的最大值，是最佳的选择。
本实施例采用的构建三角形的方式，在保证匹配准确率的前提下，可以提高匹配效率。
步骤2.3：通过学习得到先验知识参数，并将先验知识参数加入到普通目标函数中。
本实施例提出了构建与马尔科夫蒙特卡洛(MCMC)框架下的基于元算法的子超图匹配方法，为了得到更好的表现力，本实施例优选的，将先验策略融入到原始的问题(公式(4))中。
首先，获得先验知识的参数β及目标函数。本实施例在原始问题(公式(4))中加入了可靠的先验知识，也即下面的公式(5)中的第二项βTD(||x||)。其中β是通过图的学习方法来训练学习得到的，其物理意义为密实的程度，即概率密度，也是一种权重，其由相似度衡量方法和子超图的大小来决定；D(||x||)代表x的欧式距离。本实施例通过加入先验知识参数，不仅仅减小了搜索空间，更是为准确搜索和快速找到匹配对奠定了基础。
Figure PCTCN2016076030-appb-000010
下面，需要求解公式(5)，以得到最佳匹配对，作为子超图匹配结果，进行输出。
步骤2.4：构建基于MCMC框架下的目标分布函数。
定义在MCMC框架下的状态空间为Y，每一个状态y∈Y，其目标分布以及概率密度均为π(y)，因为概率密度与目标分布是成正比关系，因而用其中一个来代表即可，如下面公式(6)所示。其中，使用退火模拟值T来控制该分布的变化。f(y)为当前y状态下使得超图匹配评分Score(x)最大的值。
Figure PCTCN2016076030-appb-000011
Figure PCTCN2016076030-appb-000012
其中，Score(x)为超图匹配评分，βTD(||f(y)||)为学习得到的先验知识参数，β为概率密度，D(||f(y)||)为f(y)的欧氏距离，f(y)为当前y状态下使得超图匹配评分Score(x)最大的值，T为退火模拟值。
步骤2.5：初始化各参数，例如初始化退火模拟值T。
构建子超图匹配中的马尔科夫链，即图1中的步骤1.4，其具体包括下面子步骤。
步骤2.6：判断马尔科夫链是否达到平衡，以及退火模拟值T是否大于预设阈值Tf，如果是，则转到步骤2.7，否则转到步骤2.8。
步骤2.7：当马尔科夫链达到平衡，以及退火模拟值T大于预设阈值Tf时，说明求得的最佳匹配对已经符合要求，因此，输出匹配结果。
步骤2.8：进行模式选择(随机、增加、删除)，并计算该模式下的概率分布。
为了构建一个良好的马尔科夫链(Markov chain)，本实施例使用Metropolis-Hastings算法。状态转换核函数如下公式(7)所示，其中μ(mode)是mode的概率密度。其中mode有三个模式：随机、添加和删除。对于不同的模式，有着不同的从当前状态y到下一状态y′的概率分布q(y,y′)。需要说明的是，随机模式应用于初始阶段，随机选取第一子超图的某个点，而添加和删除模式均用于非初始阶段，也就是之后的匹配过程中，其中：μ(add)+μ(delete)＝1。μ(add)为添加模式的概率密度，μ(delete)为删除模式的概率密度。
Q(y,y′)＝μ(mode)·q(y,y′)          (7)
根据上面所说的mode模式，以及添加和删除的关系，如果假定μ(add)＝0.5，那么就意味着每次选择点的时候一半的概率是添加模式，另一半的概率是删除模式。
如图3所示，为状态转换中的添加和删除过程。其中，(a)为当前状态y的匹配关系，(b)为下一状态y′的匹配关系，(c)为删除错误的匹配对的示意图，(d)为添加新的匹配对的示意图。
不同模式下有着不同的概率分布，本实施例中，随机、添加、删除三个模式下的概率分布为(其他情况的概率分布均为0)：
Figure PCTCN2016076030-appb-000013
步骤2.9：计算当前状态y到下一状态y′的接受率α，其计算方式如公式(9)所示：
Figure PCTCN2016076030-appb-000014
步骤2.10：将计算得到的接受率α与一预设的阈值进行比较，判断接受率α是否达到阈值，以判断是否接受下一状态y′时匹配信息以及超图匹配得分，如果是则转到步骤2.11，否则转到步骤2.13。
步骤2.11：当接受率α达到预设的阈值时，再进一步判断状态转换后超图匹配评分是否提高，如果是接受状态转换，转到步骤2.12，否则转到步骤2.13。
步骤2.12：接受状态转换，更新最佳匹配对和超图匹配评分，即将子超图的最佳匹配对信息以及超图匹配得分全部都更新为y′状态下的最佳匹配对信息以及超图匹配得分。
步骤2.13：无论是否接受状态转换，在本步骤中，判断采样个数是否大于一预设阈值，如果是则转到步骤2.14，否则转到步骤2.6。
步骤2.14：更新退火模拟值T。
具体的，退火模拟值为T，采样个数为N，两者的初始值为0。采样个数对应的阈值为Nmax(即样本阈值)，退火模拟值对应的阈值为Tf(即退火模拟结束值)。当退火模拟值T大于退火模拟结束值Tf时，循环结束。对于每一次循环中，每采样一次y′,采样个数N增加1。当N>Nmax时，对T进行更新，即T＝T*C，N＝0，C在这里是退火因子。
本实施例提供的基于MCMC框架下的子超图匹配方法中，提出通过构建子超图的方式进行物体特征的匹配。由于在大量的真实图片、视频中，物体都是不断变换，包含各种噪点，以及其他一些干扰因素的，以至于图像物体匹配检索变得非常困难，而通过子超图来表现物体的外观以及位置信息来进行物体特征匹配，就能更高速、更准确地实现图像的正确匹配。并且，子超图比图、超图更有优势。一方面，子超图比图具有更多的几何信息(例如，角度变换，旋转，尺度等)，比超图的具有更低的难度，具有更好的延展性。另一方面，该方法和装置具有较强的抗干扰能力，鲁棒性好，能适应较复杂的场景，尤其是针对异常值。
本实施例提供的方法中，点之间的匹配是一个一个来进行的，而不是全部匹配完成，其通过离散化操作来真正的实现一一匹配，因此，可以使得匹配更加准确，更接近匹配所需要达到的目的。对于匹配的结果的展示，对于不断变换着的异常值(outliers)，采用召回率(recall)和精确率(precision)两个衡量方式来衡量，同时画出F曲线，F＝2*(recall*precision)/(recall+precision)，用来整体衡量匹配结果，F越高表示匹配结果越好。通过这种衡量方式来衡量匹配结果，可以很好地解 决传统衡量方法中的片面性，对于实际当中异常值是不固定的情况时更加具有实用性和普遍性。
请参考图4-9，为采用本实施例提供的方法与现有技术中的其他方法进行图像匹配时得到的召回率、准确率以及F值之间的比较示意图。
实施例二
请参考图10，基于上述实施例一提供的方法，本实施例相应提供了一种基于MCMC框架下的子超图匹配装置，包括子超图构建模块101、目标函数构建模块102、目标分布函数构建模块103和匹配结果输出模块104。
子超图构建模块101用于构建子超图，子超图包括位于原始图像中的第一子超图和位于待匹配图像中的第二子超图。
目标函数构建模块102用于构建目标函数。
目标分布函数构建模块103构建基于MCMC框架下的目标分布函数。
匹配结果输出模块104用于构建子超图匹配中的马尔科夫链，在马尔科夫链达到平衡时，根据目标函数输出子超图匹配结果。
本实施例中，匹配结果输出模块104用于在子超图匹配过程中，采用三角形张量作为元组进行匹配，并通过下面方式构建三角形张量：当第一子超图中点的数目大于第一阈值时，确定三角形的数量为两个子超图中点的数目的乘积；否则，确定三角形的数量为第一子超图中所有点能够构成的三角形的数量的最大值。
目标函数构建模块102用于通过学习得到先验知识参数，并将先验知识参数加入到目标函数中。
目标分布函数构建模块103用于定义在MCMC框架下状态空间为Y，每一个状态y∈Y，目标分布为π(y)，并且：
Figure PCTCN2016076030-appb-000015
Figure PCTCN2016076030-appb-000016
其中，Score(x)为超图匹配评分，βTD(||f(y)||)为学习得到的先验知识参数，β为概率密度，D(||f(y)||)为f(y)的欧氏距离，f(y)为当前y状态下使得超图匹配评分Score(x)最大的值，T为退火模拟值；
匹配结果输出模块用于构建子超图匹配中的马尔科夫链时，状态转换核函数为：
Q(y,y′)＝μ(mode)·q(y,y′)
其中，μ(mode)为mode的概率密度，mode包括随机、添加、删除 三个模式，q(y,y′)为从当前状态y到下一状态y′的概率分布；
随机、添加、删除三个模式下的概率分布为：
Figure PCTCN2016076030-appb-000017
本实施例提供的装置与上述实施例一提供的方法对应，此处不再对装置的具体执行步骤进行赘述。
本领域技术人员可以理解，上述实施方式中各种方法的全部或部分步骤可以通过程序来控制相关硬件完成，该程序可以存储于一计算机可读存储介质中，存储介质可以包括：只读存储器、随机存取存储器、磁盘或光盘等。
以上内容是结合具体的实施方式对本申请所作的进一步详细说明，不能认定本申请的具体实施只局限于这些说明。对于本申请所属技术领域的普通技术人员来说，在不脱离本申请发明构思的前提下，还可以做出若干简单推演或替换。

Claims (12)

  1. 一种基于MCMC框架下的子超图匹配方法，其特征在于，包括：
    构建子超图，所述子超图包括位于原始图像中的第一子超图和位于待匹配图像中的第二子超图；
    构建目标函数；
    构建基于MCMC框架下的目标分布函数；
    构建子超图匹配中的马尔科夫链，在马尔科夫链达到平衡时，根据所述目标函数输出子超图匹配结果。
  2. 如权利要求1所述的方法，其特征在于，在子超图匹配过程中，采用三角形张量作为元组进行匹配，并通过下面方式构建三角形张量：当第一子超图中点的数目大于第一阈值时，确定三角形的数量为两个子超图中点的数目的乘积；否则，确定三角形的数量为第一子超图中所有点能够构成的三角形的数量的最大值。
  3. 如权利要求1所述的方法，其特征在于，通过学习得到先验知识参数，并将所述先验知识参数加入到所述目标函数中。
  4. 如权利要求3所述的方法，其特征在于，定义在MCMC框架下状态空间为Y，每一个状态y∈Y，目标分布为π(y)，并且：
    Figure PCTCN2016076030-appb-100001
    Figure PCTCN2016076030-appb-100002
    其中，Score(x)为超图匹配评分，βTD(||f(y)||)为学习得到的先验知识参数，β为概率密度，D(||f(y)||)为f(y)的欧氏距离，f(y)为当前y状态下使得超图匹配评分Score(x)最大的值，T为退火模拟值。
  5. 如权利要求4所述的方法，其特征在于，构建子超图匹配中的马尔科夫链时，状态转换核函数为：
    Q(y,y′)＝μ(mode)·q(y,y′)
    其中，μ(mode)为mode的概率密度，mode包括随机、添加、删除三个模式，q(y,y′)为从当前状态y到下一状态y′的概率分布。
  6. 如权利要求5所述的方法，其特征在于，所述随机、添加、删除三个模式下的概率分布为：
    Figure PCTCN2016076030-appb-100003
  7. 一种基于MCMC框架下的子超图匹配装置，其特征在于，包括：
    子超图构建模块，用于构建子超图，所述子超图包括位于原始图像中的第一子超图和位于待匹配图像中的第二子超图；
    目标函数构建模块，用于构建目标函数；
    目标分布函数构建模块，构建基于MCMC框架下的目标分布函数；
    匹配结果输出模块，用于构建子超图匹配中的马尔科夫链，在马尔科夫链达到平衡时，根据所述目标函数输出子超图匹配结果。
  8. 如权利要求7所述的装置，其特征在于，匹配结果输出模块用于在子超图匹配过程中，采用三角形张量作为元组进行匹配，并通过下面方式构建三角形张量：当第一子超图中点的数目大于第一阈值时，确定三角形的数量为两个子超图中点的数目的乘积；否则，确定三角形的数量为第一子超图中所有点能够构成的三角形的数量的最大值。
  9. 如权利要求7所述的装置，其特征在于，目标函数构建模块用于通过学习得到先验知识参数，并将所述先验知识参数加入到所述目标函数中。
  10. 如权利要求9所述的装置，其特征在于，目标分布函数构建模块用于定义在MCMC框架下状态空间为Y，每一个状态y∈Y，目标分布为π(y)，并且：
    Figure PCTCN2016076030-appb-100004
    Figure PCTCN2016076030-appb-100005
    其中，Score(x)为超图匹配评分，βTD(||f(y)||)为学习得到的先验知识参数，β为概率密度，D(||f(y)||)为f(y)的欧氏距离，f(y)为当前y状态下使得超图匹配评分Score(x)最大的值，T为退火模拟值。
  11. 如权利要求10所述的装置，其特征在于，匹配结果输出模块用于构建子超图匹配中的马尔科夫链时，状态转换核函数为：
    Q(y,y′)＝μ(mode)·q(y,y′)
    其中，μ(mode)为mode的概率密度，mode包括随机、添加、删除三个模式，q(y,y′)为从当前状态y到下一状态y′的概率分布。
  12. 如权利要求11所述的装置，其特征在于，所述随机、添加、删除三个模式下的概率分布为：
    Figure PCTCN2016076030-appb-100006
PCT/CN2016/076030 2016-03-10 2016-03-10 一种基于mcmc框架下的子超图匹配方法和装置 Ceased WO2017152403A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2016/076030 WO2017152403A1 (zh) 2016-03-10 2016-03-10 一种基于mcmc框架下的子超图匹配方法和装置
US16/079,660 US11347979B2 (en) 2016-03-10 2016-03-10 MCMC framework-based sub-hypergraph matching method and device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2016/076030 WO2017152403A1 (zh) 2016-03-10 2016-03-10 一种基于mcmc框架下的子超图匹配方法和装置

Publications (1)

Publication Number Publication Date
WO2017152403A1 true WO2017152403A1 (zh) 2017-09-14

Family

ID=59789997

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2016/076030 Ceased WO2017152403A1 (zh) 2016-03-10 2016-03-10 一种基于mcmc框架下的子超图匹配方法和装置

Country Status (2)

Country Link
US (1) US11347979B2 (zh)
WO (1) WO2017152403A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107742126A (zh) * 2017-09-26 2018-02-27 南京航空航天大学 一种基于隐马尔可夫模型的浮动车地图匹配方法
CN110765854A (zh) * 2019-09-12 2020-02-07 昆明理工大学 一种视频动作识别方法
CN114676821A (zh) * 2022-03-24 2022-06-28 腾讯科技（深圳）有限公司 模型确定方法、装置、设备及计算机可读存储介质

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111967485B (zh) * 2020-04-26 2024-01-05 中国人民解放军火箭军工程大学 一种基于概率超图学习的空地红外目标跟踪方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20020095276A1 (en) * 1999-11-30 2002-07-18 Li Rong Intelligent modeling, transformation and manipulation system
CN103020120A (zh) * 2012-11-16 2013-04-03 南京理工大学 一种基于超图的图像混合摘要生成方法
CN103218598A (zh) * 2013-03-26 2013-07-24 中国科学院电子学研究所 一种基于随机几何模型的遥感地物目标自动检测方法
CN103390063A (zh) * 2013-07-31 2013-11-13 南京大学 一种基于蚁群算法和概率超图的相关反馈图像检索方法
CN105809197A (zh) * 2016-03-10 2016-07-27 北京大学深圳研究生院 一种基于mcmc框架下的子超图匹配方法和装置

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7773800B2 (en) * 2001-06-06 2010-08-10 Ying Liu Attrasoft image retrieval
EP1736928A1 (en) * 2005-06-20 2006-12-27 Mitsubishi Electric Information Technology Centre Europe B.V. Robust image registration
EP2395456A1 (en) * 2010-06-12 2011-12-14 Toyota Motor Europe NV/SA Methods and systems for semantic label propagation
EP2690596B1 (en) * 2012-07-24 2018-08-15 Agfa Healthcare Method, apparatus and system for automated spine labeling

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20020095276A1 (en) * 1999-11-30 2002-07-18 Li Rong Intelligent modeling, transformation and manipulation system
CN103020120A (zh) * 2012-11-16 2013-04-03 南京理工大学 一种基于超图的图像混合摘要生成方法
CN103218598A (zh) * 2013-03-26 2013-07-24 中国科学院电子学研究所 一种基于随机几何模型的遥感地物目标自动检测方法
CN103390063A (zh) * 2013-07-31 2013-11-13 南京大学 一种基于蚁群算法和概率超图的相关反馈图像检索方法
CN105809197A (zh) * 2016-03-10 2016-07-27 北京大学深圳研究生院 一种基于mcmc框架下的子超图匹配方法和装置

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
LIN, ET AL.: "Image Registration Algorithm Based on Sparse Position Hypergraph Matching", COMPUTER NETWORK AND MULTIMEDIA TECHNOLOGY, 2009. CNMT 2009. INTERNATIONAL SYMPOSIUM ON, vol. 21, no. 12, 18 January 2009 (2009-01-18), XP031596623 *
ZASS, R. ET AL.: "Probabilistic Graph and Hypergraph Matching", COMPUTER VISION AND PATTERN RECOGNITION , 2008 . CVPR 2008 . IEEE CONFERENCE, 5 August 2008 (2008-08-05), XP031297058, ISSN: 1063-6919 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107742126A (zh) * 2017-09-26 2018-02-27 南京航空航天大学 一种基于隐马尔可夫模型的浮动车地图匹配方法
CN110765854A (zh) * 2019-09-12 2020-02-07 昆明理工大学 一种视频动作识别方法
CN110765854B (zh) * 2019-09-12 2022-12-02 昆明理工大学 一种视频动作识别方法
CN114676821A (zh) * 2022-03-24 2022-06-28 腾讯科技（深圳）有限公司 模型确定方法、装置、设备及计算机可读存储介质

Also Published As

Publication number Publication date
US11347979B2 (en) 2022-05-31
US20190065912A1 (en) 2019-02-28

Similar Documents

Publication Publication Date Title
CN113012212B (zh) 一种基于深度信息融合的室内场景三维点云重建方法和系统
CN113971751A (zh) 训练特征提取模型、检测相似图像的方法和装置
WO2020228446A1 (zh) 模型训练方法、装置、终端及存储介质
JP2023549070A (ja) 意味特徴の学習を介したＵｎｓｅｅｎドメインからの顔認識
CN110120064A (zh) 一种基于互强化与多注意机制学习的深度相关目标跟踪算法
CN110163087A (zh) 一种人脸姿态识别方法及系统
CN110097499B (zh) 基于谱混合核高斯过程回归的单帧图像超分辨率重建方法
WO2017152403A1 (zh) 一种基于mcmc框架下的子超图匹配方法和装置
CN110956131A (zh) 单目标追踪方法、装置及系统
WO2025026318A1 (zh) 变压器的故障识别方法、装置、设备及介质
Zhang et al. Graph-regularized structured support vector machine for object tracking
Wang et al. Visual tracking via sparse and local linear coding
WO2026040716A1 (zh) 舌头图像的处理方法、装置、设备和存储介质
TWI803243B (zh) 圖像擴增方法、電腦設備及儲存介質
CN115605862A (zh) 训练用于3d模型数据库查询的可微分渲染器和神经网络
CN105809197A (zh) 一种基于mcmc框架下的子超图匹配方法和装置
CN110232365B (zh) 细胞图像中细胞轮廓弯曲程度的衡量方法、系统及介质
WO2022087778A1 (zh) 一种基于多层耦合映射的低分辨率图像识别方法
CN116524298A (zh) 基于特征存储库与对比学习的半监督目标检测方法
CN111382791A (zh) 深度学习任务处理方法、图像识别任务处理方法和装置
TWI632509B (zh) 人臉辨識裝置及方法、提升影像辨識率的方法、及電腦可讀儲存介質
CN115862055A (zh) 基于对比学习和对抗训练的行人重识别方法及装置
CN115620083B (zh) 模型训练方法、人脸图像质量评价方法、设备及介质
CN117456273A (zh) 一种尘肺ct图像分类方法、装置、计算机设备和储存介质
CN116958809B (zh) 一种特征库迁移的遥感小样本目标检测方法

Legal Events

Date Code Title Description
NENP Non-entry into the national phase

Ref country code: DE

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

Ref document number: 16893054

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 16893054

Country of ref document: EP

Kind code of ref document: A1