CN111401385B - Similarity calculation method for image local topological structure feature descriptors - Google Patents

Similarity calculation method for image local topological structure feature descriptors Download PDF

Info

Publication number
CN111401385B
CN111401385B CN202010194853.9A CN202010194853A CN111401385B CN 111401385 B CN111401385 B CN 111401385B CN 202010194853 A CN202010194853 A CN 202010194853A CN 111401385 B CN111401385 B CN 111401385B
Authority
CN
China
Prior art keywords
points
point
neighborhood
corresponding relation
distance
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.)
Expired - Fee Related
Application number
CN202010194853.9A
Other languages
Chinese (zh)
Other versions
CN111401385A (en
Inventor
桑强
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.)
Chengdu Univeristy of Technology
Original Assignee
Chengdu Univeristy of Technology
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 Chengdu Univeristy of Technology filed Critical Chengdu Univeristy of Technology
Priority to CN202010194853.9A priority Critical patent/CN111401385B/en
Publication of CN111401385A publication Critical patent/CN111401385A/en
Application granted granted Critical
Publication of CN111401385B publication Critical patent/CN111401385B/en
Expired - Fee Related legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
    • G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
    • G06V10/751—Comparing pixel values or logical combinations thereof, or feature values having positional relevance, e.g. template matching

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • Databases & Information Systems (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

本发明提出了一种图像局部拓扑结构特征描述符的相似度计算方法，首先，提出了一种局部拓扑结构特征描述符，其次，在比较局部相似度时，基于该描述符的特点，以两个偏序结构找到局部结构的对应关系，并通过提出的代价计算方法量化了不同局部拓扑结构之间的差异，从而可以有效地计算拓扑结构的相似性。本发明与现有的方法比较，精确地量化了局部结构的差异性。

Figure 202010194853

The present invention proposes a similarity calculation method for the local topological structure feature descriptor of an image. First, a local topological structure feature descriptor is proposed, and secondly, when comparing the local similarity, based on the characteristics of the descriptor, two A partial order structure is found to find the corresponding relationship of local structures, and the difference between different local topological structures is quantified by the proposed cost calculation method, so that the similarity of topological structures can be effectively calculated. Compared with the existing method, the present invention accurately quantifies the difference of the local structure.

Figure 202010194853

Description

一种图像局部拓扑结构特征描述符的相似度计算方法A Similarity Calculation Method for Image Local Topological Structure Feature Descriptors

技术领域technical field

本发明属于图像处理技术领域，尤其涉及一种图像局部拓扑结构特征描述符的相似度计算方法。The invention belongs to the technical field of image processing, and in particular relates to a similarity calculation method of a local topological structure feature descriptor of an image.

背景技术Background technique

图像的局部特征是一种图像模式，该模式度量了图像各个不同区域之间的差异性，点、边或者小的图像块都可以作为图像的局部特征。通常，局部图像定义为围绕特征中心的一小块区域，基于该区域的上下文信息生成局部描述符，再通过相应的算法来计算描述符之间的数值差，从而得到量化的特征差异。局部特征作为一种强有力的工具，已经广泛应用在计算机视觉系统和多个应用场景之中。其中，点特征由于具有良好的特征不变性和计算实时性，是最重要的一类图像特征，特别在SLAM及立体视觉算法中，能否在目标匹配过程中，保持局部稳定性，是影响到配准精度以及最终3D建模正确性的关键，这时，通过特征点提取算法，图像以离散点集合的形式表示。在计算图像局部特征时，目前主流的局部点特征计算方法是基于形状上下文(shapecontext)以及类似的基于直方图特征描述符的方法，形状上下文主要用于度量两个图像图像轮廓形状的相似性，利用了邻域点围绕特征中心点的统计分布特性，根据邻域点分布的角度和距离信息，描述了该中心点周围的形状特征。其他类似的方法有轮廓点直方图(CPDH)，以目标物体最小外接圆的圆心作为中心点，根据轮廓点的分布形成对数-极坐标直方图，在计算轮廓相似性时，使用了EMD(Earth Mover’sDistance)距离来计算两个轮廓直方图之间的相似度度量。最近很多算法，特别是3D点云配准中，在计算局部特征的时候都选取了形状上下文及类似的方法，然而，该类方法主要是用来描述图像轮廓的，而不是目标的局部结构的，形状上下文在一定程度上可以描述图像局部特征，但是，其采取的直方图方式是一种统计的方法，需要一定数量的形状轮廓点来进行投票从而形成特征直方图。而在描述目标局部结构时，由于局部性的要求，抽样点必定有限，通常为围绕中心点的几个邻居点，通常少于十个，不足以形成有效的特征直方图，其次，由于特征点过少形成的稀疏直方图也无法有效的计算局部相似性，量化相似性时，需要能够区分不同局部特征相似的程度。但在计算两个直方图之间的X2距离时，只要轮廓点分布到直方图的不同栅格上，计算得到的代价值是相同的。例如，三条角度分别为0°，45°，90°的等长边，其形状上下文计算得到的X2距离是相同的,而通常意义下，认为0°的边和45°的边更相似，而和90°的边差异更大,所以该方法无法准确体现轮廓差异的程度。The local feature of an image is an image pattern that measures the difference between different regions of the image. Points, edges or small image blocks can be used as local features of the image. Usually, a local image is defined as a small area around the feature center, a local descriptor is generated based on the context information of the area, and then the numerical difference between the descriptors is calculated by a corresponding algorithm to obtain a quantified feature difference. As a powerful tool, local features have been widely used in computer vision systems and many application scenarios. Among them, point features are the most important type of image features due to their good feature invariance and real-time calculation. Especially in SLAM and stereo vision algorithms, whether the local stability can be maintained during the target matching process is affected. The key to the registration accuracy and the final 3D modeling correctness, at this time, through the feature point extraction algorithm, the image is represented in the form of a set of discrete points. When calculating local features of an image, the current mainstream local point feature calculation methods are based on shape context and similar methods based on histogram feature descriptors. Shape context is mainly used to measure the similarity of the contour shapes of two images. Using the statistical distribution characteristics of the neighborhood points around the feature center point, according to the angle and distance information of the neighborhood point distribution, the shape features around the center point are described. Other similar methods include contour point histogram (CPDH), which takes the center of the smallest circumscribed circle of the target object as the center point, and forms a log-polar coordinate histogram according to the distribution of contour points. When calculating the contour similarity, EMD ( Earth Mover's Distance) to calculate the similarity measure between two contour histograms. Recently, many algorithms, especially in 3D point cloud registration, select shape context and similar methods when calculating local features. However, these methods are mainly used to describe the image contour, not the local structure of the target. , the shape context can describe the local features of the image to a certain extent, but the histogram method it adopts is a statistical method, which requires a certain number of shape contour points to vote to form a feature histogram. When describing the local structure of the target, due to the requirements of locality, the sampling points must be limited, usually a few neighbor points around the center point, usually less than ten, which is not enough to form an effective feature histogram. Secondly, due to the feature points The sparse histogram formed too little can not effectively calculate the local similarity. When quantifying the similarity, it is necessary to be able to distinguish the degree of similarity of different local features. But when calculating the X2 distance between two histograms, as long as the contour points are distributed on different grids of the histogram, the calculated cost value is the same. For example, for three equal-length sides with angles of 0°, 45°, and 90°, the X 2 distances calculated from the shape context are the same, and in the usual sense, the 0° side and the 45° side are considered to be more similar. The difference between the edge and 90° is larger, so this method cannot accurately reflect the degree of contour difference.

发明内容SUMMARY OF THE INVENTION

针对现有技术中的上述不足，本发明提供的一种图像局部拓扑结构特征描述符的相似度计算方法，可以有效地计算拓扑结构的相似性，与现有的方法比较，精确地量化了局部结构的差异性。Aiming at the above deficiencies in the prior art, the present invention provides a method for calculating the similarity of local topological structure feature descriptors of images, which can effectively calculate the similarity of topological structures. Compared with the existing methods, it can accurately quantify the local structural differences.

为了达到以上目的，本发明采用的技术方案为：In order to achieve the above purpose, the technical scheme adopted in the present invention is:

本方案提供一种图像局部拓扑结构特征描述符的相似度计算方法，包括以下步骤：This solution provides a similarity calculation method for image local topological structure feature descriptor, which includes the following steps:

S1、将图像轮廓线表示为点集合SP，并根据所述点集合中点pi的邻域点得到距离该中点最近的N个点，并由所述N个点构成图像的邻域结构；S1. Represent the image contour line as a point set S P , and obtain N points closest to the midpoint according to the neighborhood points of the point p i in the point set, and the N points constitute the neighborhood of the image structure;

S2、获取所述邻域点和中心点两个点之间的角度值和距离值；S2, obtain the angle value and the distance value between the two points of the neighborhood point and the center point;

S3、根据所述角度值和距离值的大小，分别得到邻域点围绕中心点的距离序列以及角度序列，并由所述距离序列以及角度序列构成图像局部拓扑结构的特征描述符；S3, according to the size of the angle value and the distance value, obtain the distance sequence and the angle sequence of the neighborhood point around the center point respectively, and the feature descriptor of the local topological structure of the image is formed by the distance sequence and the angle sequence;

S4、对比相似度时，基于所述特征描述符根据所述距离序列得到距离序列的初始对应关系，以及根据所述角度序列得到角度序列验证的对应关系；S4, when comparing the similarity, obtain the initial correspondence of the distance sequence based on the feature descriptor according to the distance sequence, and obtain the corresponding relationship verified by the angle sequence according to the angle sequence;

S5、判断所述初始对应关系与所述验证对应关系是否一致，若是，则将一致对应关系加入至最终的对应关系z中，并进入步骤S6，否则，进入步骤S7；S5, determine whether the initial corresponding relationship is consistent with the verification corresponding relationship, if so, add the consistent corresponding relationship to the final corresponding relationship z, and enter step S6, otherwise, enter step S7;

S6、根据所述最终应的对应关系z，利用代价计算方法计算得到初始对应关系和验证对应关系一致时的匹配概率P'；S6, according to the corresponding relationship z of the final response, use the cost calculation method to calculate the matching probability P' when the initial corresponding relationship and the verification corresponding relationship are consistent;

S7、利用代价计算方法计算得到初始对应关系和验证对应关系不一致时的匹配概率P”；S7, using the cost calculation method to calculate the matching probability P" when the initial corresponding relationship and the verified corresponding relationship are inconsistent;

S8、根据所述匹配概率P'和匹配概率P”量化结构差异，完成对图像局部拓扑结构特征描述符的相似度计算。S8. Quantify the structural difference according to the matching probability P' and the matching probability P", and complete the similarity calculation of the local topological structure feature descriptor of the image.

进一步地，所述步骤S3中特征描述符

Figure GDA0003561282990000031
的表达式如下：Further, the feature descriptor in the step S3
Figure GDA0003561282990000031
The expression is as follows:

Figure GDA0003561282990000032
Figure GDA0003561282990000032

Figure GDA0003561282990000033
且
Figure GDA0003561282990000034
Figure GDA0003561282990000033
and
Figure GDA0003561282990000034

Figure GDA0003561282990000035
且
Figure GDA0003561282990000036
Figure GDA0003561282990000035
and
Figure GDA0003561282990000036

其中，

Figure GDA0003561282990000037
表示距离序列，
Figure GDA0003561282990000038
表示以pi为特征中心形成的距离序列集合中的一个点，pi表示图像轮廓上的任意一个特征中心点，L(·)表示两个点之间的长度距离，
Figure GDA0003561282990000039
表示角度序列，
Figure GDA00035612829900000310
表示以pi为特征中心形成的角度序列集合中的一个点，θ(·)表示两个点之间的角度距离。in,
Figure GDA0003561282990000037
represents the distance sequence,
Figure GDA0003561282990000038
represents a point in the distance sequence set formed with pi as the feature center, pi represents any feature center point on the image contour, L( ) represents the length distance between two points,
Figure GDA0003561282990000039
represents the angle sequence,
Figure GDA00035612829900000310
represents a point in the set of angle sequences formed with pi as the feature center, and θ( ) represents the angular distance between two points.

再进一步地，所述步骤S4具体为：Further, the step S4 is specifically:

基于所述特征描述符，按照距离序列中邻域点的距离从小到大依次一一对应，得到距离序列的初始对应关系；以及Based on the feature descriptors, one-to-one correspondence is performed according to the distances of neighboring points in the distance sequence from small to large, so as to obtain the initial correspondence of the distance sequence; and

基于所述特征描述符，按照角度序列中邻域点的角度从小到大依次一一对应，得到角度序列的初始对应关系。Based on the feature descriptors, one-to-one correspondence is performed according to the angles of the neighboring points in the angle sequence from small to large, so as to obtain the initial correspondence of the angle sequence.

再进一步地，所述步骤S6中量化的结构差异Smn的表达式如下：Still further, the expression of the structural difference S mn quantified in the step S6 is as follows:

Smn＝P”+P'S mn =P"+P'

Figure GDA0003561282990000041
Figure GDA0003561282990000041

Figure GDA0003561282990000042
Figure GDA0003561282990000042

Figure GDA0003561282990000043
Figure GDA0003561282990000043

其中，P”表示初始对应关系与验证对应关系不一致时的匹配概率，P'表示初始对应关系与验证对应关系一致时的匹配概率，z表示最终的对应关系，k表示N个邻域点中有k个点匹配上了,N表示邻域点的数量，ti表示一个集合中的某点，dj表示另一个集合中的某点，x表示以ti为中心点形成邻域中的某一个点，y表示以dj为中心点形成邻域中的某一个点，tm表示和x形成邻域点对中的点，

Figure GDA0003561282990000044
表示以tm为中心点形成邻域中的点x，dn表示和x形成邻域点对中的点，
Figure GDA0003561282990000045
表示以tm为中心点形成邻域中的点y，
Figure GDA0003561282990000046
表示ti的邻域点x，
Figure GDA0003561282990000047
表示dj的邻域点y，li表示ti点对距离，lj表示dj点的对距离，lm表示计算和邻域点x组成的所有点对的距离，ln表示计算和邻域点y组成的所有点对的距离，θi表示ti的点对角度，θj表示dj点的对角度，θm表示计算和邻域点x组成的所有点对的角度，θn表示计算和邻域点y组成的所有点对的角度，
Figure GDA0003561282990000048
和
Figure GDA0003561282990000049
分别表示点对间最大的角度和距离值，C(·)表示两个点对之间的代价差。Among them, P" represents the matching probability when the initial correspondence is inconsistent with the verification correspondence, P' represents the matching probability when the initial correspondence is consistent with the verification correspondence, z represents the final correspondence, and k represents the N neighborhood points k points are matched, N represents the number of neighbor points, t i represents a point in one set, d j represents a point in another set, and x represents a certain point in the neighborhood formed by ti as the center point A point, y represents a point in the neighborhood formed by d j as the center point, t m represents the point in the neighborhood point pair formed with x,
Figure GDA0003561282990000044
Represents the point x in the neighborhood formed by t m as the center point, and d n represents the point in the neighborhood point pair formed with x,
Figure GDA0003561282990000045
represents the point y in the neighborhood formed with t m as the center point,
Figure GDA0003561282990000046
represents the neighborhood point x of ti ,
Figure GDA0003561282990000047
Represents the neighborhood point y of d j , li represents the pair distance of t i point, l j represents the pair distance of d j point, lm represents the distance of all point pairs formed by the calculation and the neighborhood point x, and ln represents the calculation sum The distance of all point pairs formed by the neighborhood point y, θ i represents the point-to-point angle of t i , θ j represents the pair angle of the d j point, θ m represents the calculated angle of all point pairs formed by the neighborhood point x, θ n represents the angle of all point pairs formed by the calculation and the neighbor point y,
Figure GDA0003561282990000048
and
Figure GDA0003561282990000049
respectively represent the maximum angle and distance between point pairs, and C( ) represents the cost difference between the two point pairs.

本发明的有益效果：Beneficial effects of the present invention:

(1)本发明提出了一个新的局部结构特征描述符，该特征描述符由特征点邻域的两个序列组成：一个序列描述邻域的距离排列关系，另一个描述了邻域的角度排列关系，这两个空间关系准确描述了局部结构的稳定性。基于该稳定性，本方法能够准确描述点局部的拓扑结构，具有良好的特征不变性，特别相较于传统直方图统计分布的方法，本描述符不受点集合稀疏性的影响，准确度高，实时性好；(1) The present invention proposes a new local structure feature descriptor, which consists of two sequences of feature point neighborhoods: one sequence describes the distance arrangement relationship of the neighborhood, and the other describes the angle arrangement of the neighborhood relationship, these two spatial relationships accurately describe the stability of the local structure. Based on this stability, this method can accurately describe the local topology of points, and has good feature invariance. Compared with the traditional histogram statistical distribution method, this descriptor is not affected by the sparsity of point sets, and has high accuracy. , the real-time performance is good;

(2)本发明提出了一种基于以上所述特征描述符的相似度计算方法。在比较局部相似度时，由描述符中两个序列描述的局部稳定关系，得到两个邻域间的对应关系，其中，根据距离序列可以得到一个初始对应关系，根据角度序列再得到一个验证对应关系，如果两个关系一致，就可以确定邻域间的一一对应关系，否则，特征不匹配，最后由该对应关系，通过提出的点对间计算方法，量化了特征的局部差异。由于该相似性计算基于局部对应关系，相似性的量化更加精确，能够准确比较图像各个不同局部区域；(2) The present invention proposes a similarity calculation method based on the above-mentioned feature descriptor. When comparing the local similarity, the local stable relationship described by the two sequences in the descriptor is used to obtain the corresponding relationship between the two neighborhoods. Among them, an initial correspondence can be obtained according to the distance sequence, and a verification correspondence can be obtained according to the angle sequence. If the two relationships are consistent, the one-to-one correspondence between the neighborhoods can be determined. Otherwise, the features do not match. Finally, the corresponding relationship is used to quantify the local difference of features through the proposed point-to-point calculation method. Since the similarity calculation is based on the local correspondence, the quantification of the similarity is more accurate, and different local regions of the image can be accurately compared;

(3)本发明的特征描述符，整合了图像特征点的上下文信息，通过邻域之间的先验对应关系，量化了图像局部特征之间的差异度，作为一种图像局部特征量化方法，在计算机视觉工程中扮演着关键的角色。(3) The feature descriptor of the present invention integrates the context information of the image feature points, and quantifies the degree of difference between the local features of the image through the prior correspondence between the neighborhoods. As a method for quantifying local features of the image, plays a key role in computer vision engineering.

附图说明Description of drawings

图1为本发明的方法流程图。FIG. 1 is a flow chart of the method of the present invention.

具体实施方式Detailed ways

下面对本发明的具体实施方式进行描述，以便于本技术领域的技术人员理解本发明，但应该清楚，本发明不限于具体实施方式的范围，对本技术领域的普通技术人员来讲，只要各种变化在所附的权利要求限定和确定的本发明的精神和范围内，这些变化是显而易见的，一切利用本发明构思的发明创造均在保护之列。The specific embodiments of the present invention are described below to facilitate those skilled in the art to understand the present invention, but it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes Such changes are obvious within the spirit and scope of the present invention as defined and determined by the appended claims, and all inventions and creations utilizing the inventive concept are within the scope of protection.

实施例Example

如图1所示，本发明提供了一种图像局部拓扑结构特征描述符的相似度计算方法，其实现方法如下：As shown in FIG. 1 , the present invention provides a method for calculating similarity of local topological structure feature descriptors of images, and the implementation method is as follows:

S1、将图像轮廓线表示为点集合SP，并根据所述点集合中点pi的邻域点得到距离该中点最近的N个点，并由所述N个点构成图像的邻域结构；S1. Represent the image contour line as a point set S P , and obtain N points closest to the midpoint according to the neighborhood points of the point p i in the point set, and the N points constitute the neighborhood of the image structure;

S2、获取所述邻域点和中心点两个点之间的角度值和距离值；S2, obtain the angle value and the distance value between the two points of the neighborhood point and the center point;

S3、根据所述角度值和距离值的大小，分别得到邻域点围绕中心点的距离序列以及角度序列，并由所述距离序列以及角度序列构成图像局部拓扑结构的特征描述符；S3, according to the size of the angle value and the distance value, obtain the distance sequence and the angle sequence of the neighborhood point around the center point respectively, and the feature descriptor of the local topological structure of the image is formed by the distance sequence and the angle sequence;

S4、对比相似度时，基于所述特征描述符根据所述距离序列得到距离序列的初始对应关系，以及根据所述角度序列得到角度序列验证的对应关系；S4, when comparing the similarity, obtain the initial correspondence of the distance sequence based on the feature descriptor according to the distance sequence, and obtain the corresponding relationship verified by the angle sequence according to the angle sequence;

S5、判断所述初始对应关系与所述验证对应关系是否一致，若是，则将一致对应关系加入至最终的对应关系z中，并进入步骤S6，否则，进入步骤S7；S5, determine whether the initial corresponding relationship is consistent with the verification corresponding relationship, if so, add the consistent corresponding relationship to the final corresponding relationship z, and enter step S6, otherwise, enter step S7;

S6、根据所述最终应的对应关系z，利用代价计算方法计算得到初始对应关系和验证对应关系一致时的匹配概率P'；S6, according to the corresponding relationship z of the final response, use the cost calculation method to calculate the matching probability P' when the initial corresponding relationship and the verification corresponding relationship are consistent;

S7、利用代价计算方法计算得到初始对应关系和验证对应关系不一致时的匹配概率P”；S7, using the cost calculation method to calculate the matching probability P" when the initial corresponding relationship and the verified corresponding relationship are inconsistent;

S8、根据所述匹配概率P'和匹配概率P”量化结构差异，完成对图像局部拓扑结构特征描述符的相似度计算。S8. Quantify the structural difference according to the matching probability P' and the matching probability P", and complete the similarity calculation of the local topological structure feature descriptor of the image.

本实施例中，首先一个图像由其轮廓线上的抽样点组成，那么一个轮廓可以表示为点集合:SP＝p1,p2…,pM,其中,M为点集合的势。集合中点pi的邻域为

Figure GDA0003561282990000071
定义为距离该点最近的N个点，其中,o＝1,2,...,N。该点的局部拓扑结构由这N个点所构成，并且能够成为一个稳定的结构，定义邻居点和中心点的关系由这两个点之间的角度值和距离值构成，那么这N个点对之间的关系总和就唯一地确定了中心点的局部拓扑结构，按照角度值和距离值的大小，邻域点围绕中心点可以形成两个有序集合。
Figure GDA0003561282990000072
其中，满足
Figure GDA0003561282990000073
L(·)表示两个点之间的长度距离，这里使用了L2距离。同理，
Figure GDA0003561282990000074
其中，满足
Figure GDA0003561282990000075
θ(·)表示两个点之间的角度距离。本发明提出的新的局部结构描述符由这两个有序集合构成。对于点pi该描述符
Figure GDA0003561282990000076
In this embodiment, first, an image is composed of sampling points on its contour line, then a contour can be represented as a point set: S P = p 1 , p 2 . . . , p M , where M is the potential of the point set. The neighborhood of point p i in the set is
Figure GDA0003561282990000071
Defined as the N points closest to the point, where o=1,2,...,N. The local topological structure of the point is composed of these N points, and it can become a stable structure. The relationship between the neighbor point and the center point is defined by the angle value and distance value between these two points, then the N points The sum of the relationship between the pairs uniquely determines the local topology of the center point. According to the size of the angle value and the distance value, the neighborhood points can form two ordered sets around the center point.
Figure GDA0003561282990000072
Among them, satisfying
Figure GDA0003561282990000073
L(·) represents the length distance between two points, here L 2 distance is used. Similarly,
Figure GDA0003561282990000074
Among them, satisfying
Figure GDA0003561282990000075
θ(·) represents the angular distance between two points. The new local structure descriptor proposed by the present invention consists of these two ordered sets. For point pi the descriptor
Figure GDA0003561282990000076

本实施例中，当该轮廓点集合发生刚性形变时，由于点对之间的距离和角度关系没有发生变化，该描述符保持一致。当轮廓点发生弹性形变时，点集合发生的是各向异性变化，即点对之间的距离和角度值的变化是任意的。然而，对于在工程应用中遇到的目标物体发生的弹性形变，必定是具有一定意义的。例如，呼吸时内脏器官的形变、动物行走时外形轮廓的形变和不同手写体汉字的外形轮廓的变化。虽然这些不同目标的外形变化都是弹性形变，但是却遵循了一定的规律，就是该目标在发生了弹性形变后，所包含的内在意义没有变化,即，心脏在呼吸变化间虽然外形发生变化，但是其轮廓特征仍然表示其属于一个心脏器官,或者行进间的一头猎豹，不断变化的外形特征仍可以使人们一眼认出这是一头猎豹。如果变形是任意的，目标已经完全无法辨认，就失去了后续的应用价值，毫无意义。所以，基于微分的思想,在整体上，图像轮廓发生的是弹性形变，即各向异性的变化；在局部，点之间的变化是相对平滑的，近似于各向同性的变化,点和其邻域点之间角度和长度的变化是相一致的。即邻域点在中心点周围的排列次序是稳定的,在整体轮廓发生弹性形变后，局部点Tpi的拓扑描述符中的这两个有序集合是稳定不变的。In this embodiment, when the set of contour points undergoes rigid deformation, since the distance and angular relationship between the point pairs do not change, the descriptor remains consistent. When the contour points are elastically deformed, the anisotropic change occurs in the point set, that is, the changes in the distance and angle values between the point pairs are arbitrary. However, the elastic deformation of the target object encountered in engineering applications must be meaningful. For example, the deformation of internal organs when breathing, the deformation of the outline of animals when walking, and the changes of the outline of different handwritten Chinese characters. Although the shape changes of these different targets are elastic deformation, they follow a certain rule, that is, after the elastic deformation of the target, the inner meaning contained in it does not change, that is, although the shape of the heart changes during the breathing change, But its outline features still indicate that it belongs to a heart organ, or a cheetah in the process, and the changing shape features can still make people recognize it as a cheetah at a glance. If the deformation is arbitrary and the target is completely unrecognizable, it loses its subsequent application value and is meaningless. Therefore, based on the idea of differentiation, on the whole, the image outline is elastically deformed, that is, anisotropic change; locally, the change between points is relatively smooth, similar to the isotropic change, the point and its Changes in angle and length between neighboring points are consistent. That is, the arrangement order of the neighborhood points around the center point is stable, and after the elastic deformation of the overall outline, the two ordered sets in the topological descriptor of the local point T pi are stable and invariant.

本实施例中，计算两个局部结构相似性时，如果邻域点的分布完全相同，我们可以认为这两个局部结构相似度最大，反之，相似度最小。当邻域点数量等于一时(N＝1)，只需计算两个点对之间的角度差和距离差，就可以得到局部结构代价差，因为仅有的一个邻域点即互为对应点，而当N>＝2时，关键点在于如何确定局部结构邻域点间的对应关系。本发明根据以上描述的拓扑描述符，解决了局部结构之间的对应关系，因为描述符中两个集合的点序列在整个点集合发生弹性形变时，局部邻域点序列保持稳定，所以，对于两个未知的局部结构，相应的两个描述符

Figure GDA0003561282990000081
和
Figure GDA0003561282990000082
确定的对应点对序列同样稳定。首先，根据描述符
Figure GDA0003561282990000083
中的
Figure GDA0003561282990000084
和
Figure GDA0003561282990000085
中的
Figure GDA0003561282990000086
可以得到根据距离值的大小进行排序得到的两个邻域点集之间的对应关系。其次，根据
Figure GDA0003561282990000087
中的
Figure GDA0003561282990000088
和
Figure GDA0003561282990000089
中的
Figure GDA00035612829900000810
可以得到根据角度值大小排序的先后次序对应得到的两个邻域点集之间的对应关系，即，按照邻域点距离和角度值从小到大的次序一一对应，有
Figure GDA00035612829900000811
对应于点
Figure GDA00035612829900000812
或有
Figure GDA00035612829900000813
对应于点
Figure GDA00035612829900000814
并将得到的这两个对应关系分别记为：zl和zθ,如果两个点在zl和zθ中的对应关系是一致的，那么将这两个点标记为对应点,如果两局部结构中的所有点都满足上述对应关系，那么可以认为这两个局部结构是完全相似的，否则，不相似或者部分相似。设局部结构中有k个点是互相对应的,将该对应关系计为z。同时有N-k个点是没有匹配的，对于匹配的点对,给出量化的局部结构相似度概率为：In this embodiment, when calculating the similarity of two local structures, if the distribution of the neighborhood points is exactly the same, we can consider that the similarity of the two local structures is the largest, otherwise, the similarity is the smallest. When the number of neighbor points is equal to one (N=1), the local structural cost difference can be obtained by simply calculating the angle difference and distance difference between the two point pairs, because only one neighbor point corresponds to each other. , and when N>=2, the key point is how to determine the correspondence between the local structure neighborhood points. According to the topological descriptor described above, the present invention solves the correspondence between the local structures, because the point sequence of the two sets in the descriptor keeps stable when the whole point set elastically deforms, so, for Two unknown local structures, corresponding to two descriptors
Figure GDA0003561282990000081
and
Figure GDA0003561282990000082
The determined corresponding points are also stable to the sequence. First, according to the descriptor
Figure GDA0003561282990000083
middle
Figure GDA0003561282990000084
and
Figure GDA0003561282990000085
middle
Figure GDA0003561282990000086
The correspondence between the two neighborhood point sets obtained by sorting according to the size of the distance value can be obtained. Second, according to
Figure GDA0003561282990000087
middle
Figure GDA0003561282990000088
and
Figure GDA0003561282990000089
middle
Figure GDA00035612829900000810
The corresponding relationship between the two neighborhood point sets obtained according to the order of the angle values can be obtained, that is, according to the distance of the neighborhood points and the order of the angle values from small to large, there is a one-to-one correspondence.
Figure GDA00035612829900000811
corresponds to the point
Figure GDA00035612829900000812
might have
Figure GDA00035612829900000813
corresponds to the point
Figure GDA00035612829900000814
And the two obtained correspondences are recorded as: z l and z θ , if the correspondence between the two points in z l and z θ is the same, then mark these two points as corresponding points, if the two If all points in the local structure satisfy the above correspondence, then the two local structures can be considered to be completely similar, otherwise, they are not similar or partially similar. Suppose there are k points in the local structure that correspond to each other, and the correspondence is counted as z. At the same time, there are Nk points that are not matched. For the matched point pairs, the quantified local structure similarity probability is given as:

Figure GDA00035612829900000815
Figure GDA00035612829900000815

对于为未匹配上的点对,因为缺乏相应的对应信息,所以匹配概率为：For the point pairs that are not matched, because of the lack of corresponding corresponding information, the matching probability is:

Figure GDA0003561282990000091
Figure GDA0003561282990000091

这里,C(·)为两个点对之间的代价差：Here, C( ) is the cost difference between two pairs of points:

Figure GDA0003561282990000092
Figure GDA0003561282990000092

上式中，

Figure GDA0003561282990000093
和
Figure GDA0003561282990000094
分别表示点对间最大的距离和角度值，最终的相似度计算概率为：Smn＝P”+P'。In the above formula,
Figure GDA0003561282990000093
and
Figure GDA0003561282990000094
respectively represent the maximum distance and angle value between point pairs, and the final similarity calculation probability is: S mn =P"+P'.

综上所述，本发明提出了一个新的局部结构描述符，该描述符由特征点邻域的两个序列组成：一个序列描述邻域的距离排列关系，另一个描述了邻域的角度排列关系，这两个空间关系准确描述了局部结构的稳定性，其次，基于这两个序列描述的局部稳定关系，可以得到两个邻域间的对应关系，其中，根据距离序列可以得到一个初始对应关系，根据角度序列再得到一个验证对应关系，如果两个关系一致，就得到了邻域间的对应关系。否则，特征不匹配，最后由得到的对应关系，计算量化的结构差异，本发明与之前的方法比较，能精确地量化局部结构的差异性。To sum up, the present invention proposes a new local structure descriptor, which consists of two sequences of feature point neighborhoods: one sequence describes the distance arrangement relationship of the neighborhood, and the other describes the angle arrangement of the neighborhood The two spatial relationships accurately describe the stability of the local structure. Secondly, based on the local stability relationship described by the two sequences, the corresponding relationship between the two neighborhoods can be obtained. An initial correspondence can be obtained according to the distance sequence. According to the angle sequence, a verification corresponding relationship is obtained. If the two relationships are consistent, the corresponding relationship between the neighborhoods is obtained. Otherwise, the features do not match, and finally the quantified structural difference is calculated from the obtained correspondence. Compared with the previous method, the present invention can accurately quantify the local structural difference.

Claims (4)

1. A similarity calculation method for image local topological structure feature descriptors is characterized by comprising the following steps:
s1, representing the image contour as a point set SPAnd according to said points, collecting points piObtaining N points nearest to the midpoint from the neighborhood points, and forming a neighborhood structure of the image by the N points;
s2, obtaining an angle value and a distance value between the two points of the neighborhood point and the central point;
s3, respectively obtaining a distance sequence and an angle sequence of the neighborhood points around the central point according to the angle value and the distance value, and forming a feature descriptor of the image local topological structure by the distance sequence and the angle sequence;
s4, when comparing the similarity, based on the feature descriptor, obtaining the initial corresponding relation of the distance sequence according to the distance sequence, and obtaining the verification corresponding relation of the angle sequence according to the angle sequence;
s5, judging whether the initial corresponding relation is consistent with the verification corresponding relation, if so, adding the consistent corresponding relation into the final corresponding relation z, and entering the step S6, otherwise, entering the step S7;
s6, calculating by using a cost calculation method according to the final corresponding relation z to obtain a matching probability P' when the initial corresponding relation is consistent with the verification corresponding relation;
s7, calculating by using a cost calculation method to obtain a matching probability P' when the initial corresponding relation is inconsistent with the verification corresponding relation;
s8, according to the matching probability P 'and the matching probability P' quantification structure difference, completing the similarity calculation of the image local topological structure feature descriptor.
2. The method for calculating similarity of image local topological structure feature descriptors according to claim 1, wherein said feature descriptors in step S3
Figure FDA0003561282980000011
The expression of (c) is as follows:
Figure FDA0003561282980000012
Figure FDA0003561282980000013
and is provided with
Figure FDA0003561282980000014
Figure FDA0003561282980000015
And is
Figure FDA0003561282980000016
Wherein,
Figure FDA0003561282980000021
a sequence of distances is represented which is,
Figure FDA0003561282980000022
is represented by piOne point of the set of distance sequences formed for the feature center, piRepresents any feature center point on the image contour, L (-) represents the length distance between two points,
Figure FDA0003561282980000023
a sequence of angles is represented which is,
Figure FDA0003561282980000024
is represented by piFor one point in the set of angular sequences formed for the feature center, θ (-) represents the angular distance between the two points.
3. The method for calculating the similarity of the image local topological structure feature descriptors according to claim 1, wherein the step S4 specifically includes:
based on the feature descriptors, sequentially corresponding the distances of the neighborhood points in the distance sequence from small to large one by one to obtain an initial corresponding relation of the distance sequence; and
and based on the feature descriptors, sequentially corresponding the angles of the neighborhood points in the angle sequence from small to large one by one to obtain the initial corresponding relation of the angle sequence.
4. The method for calculating similarity of image local topological structure feature descriptors according to claim 1, wherein said structure difference S quantified in step S8 is SmnThe expression of (a) is as follows:
Smn＝P”+P'
Figure FDA0003561282980000025
Figure FDA0003561282980000026
Figure FDA0003561282980000027
wherein, P 'represents the matching probability when the initial corresponding relation is inconsistent with the verification corresponding relation, P' represents the matching probability when the initial corresponding relation is consistent with the verification corresponding relation, z represents the final corresponding relation, k represents that k points in N neighborhood points are matched, N represents the number of the neighborhood points, t represents the matching probability when the initial corresponding relation is inconsistent with the verification corresponding relation, k represents the matching probability when the initial corresponding relation is consistent with the verification corresponding relation, k represents that k points in the N neighborhood points are matched, andirepresenting in a collectionA certain point, djRepresenting a point in another set, x being denoted by tiForming a certain point in the neighborhood for the center point, y being denoted by djForming a certain point in the neighborhood for the central point, tmThe representation and x form the points in the neighborhood point pair,
Figure FDA0003561282980000031
is represented by tmForming points x, d in the neighborhood for the center pointnThe representation and x form the points in the neighborhood point pair,
Figure FDA0003561282980000032
is represented by tmA point y in the neighborhood is formed for the center point,
Figure FDA0003561282980000033
denotes tiThe neighborhood of points x of the image data,
Figure FDA0003561282980000034
denotes djNeighborhood point y, liRepresents tiPoint to point distance,/jDenotes djPair distance of points,/mRepresenting the distance, l, of all pairs of points computed and composed of the neighborhood points xnRepresenting the distance, θ, of all pairs of points computed and made up of the neighborhood points yiRepresents tiPoint to angle of (theta)jDenotes djDiagonal of the point, θmRepresenting the angle, θ, of all pairs of points computed and made up of the neighborhood points xnRepresenting the angles of all pairs of points computed and made up of the neighborhood points y,
Figure FDA0003561282980000035
and
Figure FDA0003561282980000036
respectively, the maximum angle and distance values between the pairs of points, and C (-) represents the cost difference between the two pairs of points.
CN202010194853.9A 2020-03-19 2020-03-19 Similarity calculation method for image local topological structure feature descriptors Expired - Fee Related CN111401385B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010194853.9A CN111401385B (en) 2020-03-19 2020-03-19 Similarity calculation method for image local topological structure feature descriptors

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010194853.9A CN111401385B (en) 2020-03-19 2020-03-19 Similarity calculation method for image local topological structure feature descriptors

Publications (2)

Publication Number Publication Date
CN111401385A CN111401385A (en) 2020-07-10
CN111401385B true CN111401385B (en) 2022-06-17

Family

ID=71436972

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010194853.9A Expired - Fee Related CN111401385B (en) 2020-03-19 2020-03-19 Similarity calculation method for image local topological structure feature descriptors

Country Status (1)

Country Link
CN (1) CN111401385B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117235545B (en) * 2023-11-14 2024-02-02 亿海蓝（北京）数据技术股份公司 Method, device and readable storage medium for judging spatial object distribution similarity

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8520949B1 (en) * 2008-06-20 2013-08-27 Google Inc. Self-similar descriptor filtering
CN104835116A (en) * 2015-04-10 2015-08-12 山东师范大学 Contour-based two-dimensional fragment splicing method
CN105405116A (en) * 2014-09-04 2016-03-16 南京理工大学 Stereo matching method based on graph cut
CN106530338A (en) * 2016-10-31 2017-03-22 武汉纺织大学 Method and system for matching MR image feature points before and after nonlinear deformation of biological tissue
CN106530341A (en) * 2016-11-01 2017-03-22 成都理工大学 Point registration algorithm capable of keeping local topology invariance
CN107808362A (en) * 2017-11-15 2018-03-16 北京工业大学 A kind of image split-joint method combined based on unmanned plane POS information with image SURF features
CN108416732A (en) * 2018-02-02 2018-08-17 重庆邮电大学 A Panoramic Image Stitching Method Based on Image Registration and Multi-resolution Fusion
CN109472816A (en) * 2018-09-17 2019-03-15 西北大学 A point cloud registration method
CN109948624A (en) * 2019-02-18 2019-06-28 北京旷视科技有限公司 Method, apparatus, electronic equipment and the computer storage medium of feature extraction

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6996254B2 (en) * 2001-06-18 2006-02-07 Microsoft Corporation Incremental motion estimation through local bundle adjustment
FR2983607B1 (en) * 2011-12-02 2014-01-17 Morpho METHOD AND DEVICE FOR TRACKING AN OBJECT IN A SEQUENCE OF AT LEAST TWO IMAGES

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8520949B1 (en) * 2008-06-20 2013-08-27 Google Inc. Self-similar descriptor filtering
CN105405116A (en) * 2014-09-04 2016-03-16 南京理工大学 Stereo matching method based on graph cut
CN104835116A (en) * 2015-04-10 2015-08-12 山东师范大学 Contour-based two-dimensional fragment splicing method
CN106530338A (en) * 2016-10-31 2017-03-22 武汉纺织大学 Method and system for matching MR image feature points before and after nonlinear deformation of biological tissue
CN106530341A (en) * 2016-11-01 2017-03-22 成都理工大学 Point registration algorithm capable of keeping local topology invariance
CN107808362A (en) * 2017-11-15 2018-03-16 北京工业大学 A kind of image split-joint method combined based on unmanned plane POS information with image SURF features
CN108416732A (en) * 2018-02-02 2018-08-17 重庆邮电大学 A Panoramic Image Stitching Method Based on Image Registration and Multi-resolution Fusion
CN109472816A (en) * 2018-09-17 2019-03-15 西北大学 A point cloud registration method
CN109948624A (en) * 2019-02-18 2019-06-28 北京旷视科技有限公司 Method, apparatus, electronic equipment and the computer storage medium of feature extraction

Non-Patent Citations (6)

* Cited by examiner, † Cited by third party
Title
A Statistical Approach to the Matching of Local Features;J. Rabin et al;《SIAM Journal on Imaging Sciences》;20090918;第02卷(第03期);931-958 *
Application of neighborhood feature in point clouds registration;Yong-Xing He et al;《 Proceedings of 2011 International Conference on Computer Science and Network Technology》;20120412;838-843 *
Information Theoretic Shape Matching;Erion Hasanbelliu et al;《 IEEE Transactions on Pattern Analysis and Machine Intelligence》;20140516;第36卷(第12期);2436 - 2451 *
Robust non-rigid point registration based on feature-dependant finite mixture model;Qiang Sang et al;《PATTERN RECOGNITION LETTERS》;20130911;第34卷(第13期);1557-1565 *
图像特征点匹配关键技术研究;翟凤文;《中国博士学位论文全文数据库 信息科技辑》;20200315(第03期);I138-47 *
复杂场景图像特征分析与匹配方法研究;杨燕;《中国博士学位论文全文数据库 信息科技辑》;20200115(第01期);I138-115 *

Also Published As

Publication number Publication date
CN111401385A (en) 2020-07-10

Similar Documents

Publication Publication Date Title
CN109887015B (en) A Point Cloud Automatic Registration Method Based on Local Surface Feature Histogram
WO2024077812A1 (en) Single building three-dimensional reconstruction method based on point cloud semantic segmentation and structure fitting
CN109697692B (en) Feature matching method based on local structure similarity
CN106355577B (en) Fast Image Matching Method and System Based on Feature State and Global Consistency
CN105740798B (en) A kind of point cloud object scene recognition methods based on structural analysis
CN107944020A (en) Facial image lookup method and device, computer installation and storage medium
CN107346550B (en) A fast registration method for 3D point cloud data with color information
CN104765768A (en) Mass face database rapid and accurate retrieval method
CN106530341B (en) A Point Registration Algorithm Preserving Local Topological Invariance
CN113409332B (en) Building plane segmentation method based on three-dimensional point cloud
CN105809672A (en) Super pixels and structure constraint based image's multiple targets synchronous segmentation method
CN109934298B (en) Progressive graph matching method and device of deformation graph based on clustering
Liu et al. A novel rock-mass point cloud registration method based on feature line extraction and feature point matching
CN103955950B (en) Image tracking method utilizing key point feature matching
CN112102382A (en) Electromechanical equipment visual information stereo matching algorithm based on multi-scale transformation and ADcensus-JWGF
TW202217662A (en) Visual positioning method, training method of related models, electronic device and computer-readable storage medium
CN105279508A (en) Medical image classification method based on KAP digraph model
CN107451617B (en) A Graph Transduced Semi-Supervised Classification Method
CN108764157A (en) Building laser footpoint extracting method and system based on normal vector Gaussian Profile
CN114170450A (en) An Image Feature Matching Method Based on Regional Consistency and GMS
CN117541807A (en) Terrain characteristic line extraction method based on active learning
CN111401385B (en) Similarity calculation method for image local topological structure feature descriptors
CN109887012B (en) Point cloud registration method combined with self-adaptive search point set
CN113343765B (en) Scene retrieval method and system based on point cloud rigid registration
CN114565774A (en) 3D (three-dimensional) image volume integral classification method based on local geometry and global structure joint learning

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
CF01 Termination of patent right due to non-payment of annual fee

Granted publication date: 20220617

CF01 Termination of patent right due to non-payment of annual fee