CN102819836A

Movatterモバイル変換

Info

Publication number: CN102819836A
Application number: CN2012102243583A
Authority: CN
Inventors: 王晓茹; 余志洪; 邬书哲; 李旭; 辛海明; 张宇
Original assignee: Beijing University of Posts and Telecommunications
Current assignee: Beijing University of Posts and Telecommunications
Priority date: 2012-06-28
Filing date: 2012-06-28
Publication date: 2012-12-12
Anticipated expiration: 2032-06-28
Also published as: CN102819836B

Abstract

本发明实施例提供一种图像分割方法及系统。该方法包括对给定图像进行预分割，获得多个子区域；按照至少两种特征分别对所述多个子区域进行合并，获得每种特征对应的子分割结果；将所述至少两种特征对应的子分割结果表示为超图；基于所述超图将所述至少两种特征的子分割结果进行聚类集成，获得所述给定图像的分割结果。该分割方法有效地结合了多种特征对图像的分割，克服了一般分割算法因为基于单一的图像特征而不具备通用性的缺点，同时，避免了因为视觉特征向量和Bag of words中各部分权值和维数的设置不当而使得特征对分割的性能无法起到应有作用的问题。

Embodiments of the present invention provide an image segmentation method and system. The method includes pre-segmenting a given image to obtain a plurality of sub-regions; respectively merging the plurality of sub-regions according to at least two features to obtain a sub-segmentation result corresponding to each feature; The sub-segmentation results are represented as a hypergraph; the sub-segmentation results of the at least two features are clustered and integrated based on the hypergraph to obtain the segmentation results of the given image. This segmentation method effectively combines a variety of features to segment images, overcomes the shortcomings of general segmentation algorithms that are not universal because they are based on a single image feature, and at the same time, avoids the problem of visual feature vectors and weights of each part in Bag of words. The improper setting of the value and dimension makes the feature unable to play its due role in the performance of the segmentation.

Description

Translated fromChinese

一种图像分割方法及系统A method and system for image segmentation

技术领域technical field

本发明涉及图像处理技术领域，尤其涉及一种图像分割方法及系统。The present invention relates to the technical field of image processing, in particular to an image segmentation method and system.

背景技术Background technique

图像分割是图像处理和分析的第一步，是图像目标识别的核心技术，同时也是图像处理中最古老和最困难的问题之一。图像分割是将图像表示为物理上有意义的连通区域的集合，也就是根据目标与背景的先验知识，对图像中的目标、背景进行标记、定位，然后将目标从背景或其他伪目标中分离出来的过程。Image segmentation is the first step in image processing and analysis, the core technology of image target recognition, and one of the oldest and most difficult problems in image processing. Image segmentation is to represent an image as a collection of physically meaningful connected regions, that is, to mark and locate the target and background in the image according to the prior knowledge of the target and the background, and then separate the target from the background or other pseudo targets. process of separation.

图像分割的方法有多种，任何基于单一特征的分割算法都只能在对该特征敏感的图像类上产生较好的分割效果，而对于该特征不适应的图像则分割的性能下降很大，因此，融合了多种特征的分割算法才能对大多数的图像均获得较为理想的分割性能。There are many methods for image segmentation. Any segmentation algorithm based on a single feature can only produce better segmentation results on image classes that are sensitive to this feature, while the performance of segmentation for images that are not suitable for this feature will drop greatly. Therefore, a segmentation algorithm that combines multiple features can obtain ideal segmentation performance for most images.

现有技术中，采用将提取的多个全局特征组成一个高维的特征向量的方式来融合多特征的分割方法，由于每种视觉特征所占的维度不同，因此维度高的特征往往在分割中起到了主导作用，而其它特征难以发挥作用，分割的性能难以提高；采用基于词袋（Bag of words）的方式将提取的多个全局特征和局部特征组成视觉单词来融合多特征的分割方式，由于全局特征和局部特征性质不同，形成的视觉词也是明显不同的，因此，简单地将这两种视觉词直接在分割中联合起来使用并不能发挥每种特征的作用，分割的性能难以改善。由于视觉特征向量和Bag of words中各部分维数和权值的设置不当，使得特征对分割的性能无法起到应有的作用，因此，这些方式无法发挥每种特征在分割中起到的作用。In the prior art, the multi-feature segmentation method is combined by combining multiple extracted global features into a high-dimensional feature vector. Since each visual feature occupies a different dimension, the high-dimensional features are often in the segmentation process. It plays a leading role, but other features are difficult to play a role, and the performance of segmentation is difficult to improve; the method based on bag of words (Bag of words) is used to combine multiple extracted global features and local features into visual words to fuse multiple features. Due to the different properties of global features and local features, the formed visual words are also obviously different. Therefore, simply combining these two visual words directly in segmentation cannot play the role of each feature, and the performance of segmentation is difficult to improve. Due to the improper setting of the dimensions and weights of each part in the visual feature vector and Bag of words, the features cannot play their due role in the performance of the segmentation. Therefore, these methods cannot play the role of each feature in the segmentation. .

发明内容Contents of the invention

本发明实施例提供一种图像分割方法及系统，能够更有效的融合多种特征的分割结果。Embodiments of the present invention provide an image segmentation method and system, which can more effectively fuse segmentation results of various features.

为了解决上述技术问题，本发明实施例的技术方案如下：In order to solve the above technical problems, the technical solutions of the embodiments of the present invention are as follows:

一种图像分割方法，包括：A method for image segmentation, comprising:

对给定图像进行预分割，获得多个子区域；Pre-segment a given image to obtain multiple sub-regions;

按照至少两种特征分别对所述多个子区域进行合并，获得每种特征对应的子分割结果；Merge the plurality of sub-regions according to at least two features to obtain a sub-segmentation result corresponding to each feature;

将所述至少两种特征对应的子分割结果表示为超图；Representing the sub-segmentation results corresponding to the at least two features as a hypergraph;

基于所述超图将所述至少两种特征的子分割结果进行聚类集成，获得所述给定图像的分割结果。The sub-segmentation results of the at least two features are clustered and integrated based on the hypergraph to obtain the segmentation result of the given image.

进一步，所述对所述多个子区域进行合并，包括：Further, the merging of the multiple sub-regions includes:

确定每个子区域的链接区域；determine the link area for each sub-area;

计算所述每个子区域的链接区域的合并权值；calculating the combined weight of the linked areas of each sub-area;

将所述每个子区域与其合并权值满足条件的链接区域进行合并。Merge each sub-area with its link area whose merge weight meets the conditions.

进一步，所述确定每个子区域的链接区域，包括：Further, said determining the link area of each sub-area includes:

提取子区域与其邻接子区域的特征值；Extract the eigenvalues of the subregion and its adjacent subregions;

根据提取的特征值，计算所述子区域与其邻接子区域之间的特征相似度；Calculate the feature similarity between the sub-region and its adjacent sub-regions according to the extracted feature value;

将特征相似度大于相似度阈值，且区域面积大于所述子区域的邻接子区域作为所述子区域的链接区域。Adjacent sub-regions whose feature similarity is greater than a similarity threshold and whose area is larger than the sub-region are used as link regions of the sub-region.

进一步，所述计算所述每个子区域的链接区域的合并权值，包括：Further, the calculation of the combined weight of the link areas of each sub-area includes:

根据所述每个子区域与其链接区域间的特征相似度及语义相似度评估函数，按照网页级别算法，确定所述每个子区域的链接区域的合并权值。According to the feature similarity and semantic similarity evaluation function between each sub-area and its link area, and according to the webpage level algorithm, determine the merging weight of the link area of each sub-area.

进一步，所述语义相似度评估函数为正态分布函数或折线函数。Further, the semantic similarity evaluation function is a normal distribution function or a broken line function.

进一步，所述将所述每个子区域与其合并权值满足条件的链接区域进行合并，包括：Further, the merging of each sub-region and its link region whose merging weight meets the conditions includes:

将所述每个子区域与其合并权值最大的链接区域进行合并。Merge each sub-area with the link area with the largest merging weight.

进一步，所述基于所述超图将所述至少两种特征的子分割结果进行聚类集成，获得所述给定图像的分割结果，包括：Further, the clustering and integration of the sub-segmentation results of the at least two features based on the hypergraph to obtain the segmentation result of the given image includes:

设定最终聚类数目；Set the final number of clusters;

基于所述超图将所述至少两种特征的子分割结果按照所述最终聚类数目进行谱聚类集成，获得所述给定图像的分割结果。performing spectral clustering integration on the sub-segmentation results of the at least two features according to the final number of clusters based on the hypergraph to obtain a segmentation result of the given image.

进一步，所述设定最终聚类数目包括：Further, the setting of the final number of clusters includes:

针对所述至少两种特征对应的子分割结果分别计算特征相似度；Calculating feature similarities for the sub-segmentation results corresponding to the at least two features;

对计算获得的特征相似度进行归一化；Normalize the calculated feature similarity;

将归一化后特征相似度的最小值所对应的子分割结果的聚类数目作为所述最终聚类数目。The number of clusters of the sub-segmentation results corresponding to the minimum value of the feature similarity after normalization is taken as the final number of clusters.

一种图像分割系统，包括：An image segmentation system comprising:

预分割单元，用于对给定图像进行预分割，获得多个子区域；A pre-segmentation unit, configured to pre-segment a given image to obtain multiple sub-regions;

合并单元，用于按照至少两种特征分别对所述多个子区域进行合并，获得每种特征对应的子分割结果；A merging unit, configured to merge the plurality of sub-regions according to at least two features, to obtain a sub-segmentation result corresponding to each feature;

转换单元，用于将所述至少两种特征对应的子分割结果表示为超图；A conversion unit, configured to represent the sub-segmentation results corresponding to the at least two features as a hypergraph;

分割单元，用于基于所述超图将所述至少两种特征的子分割结果进行聚类集成，获得所述给定图像的分割结果。A segmentation unit, configured to cluster and integrate the sub-segmentation results of the at least two features based on the hypergraph, to obtain the segmentation result of the given image.

进一步，所述合并单元包括：Further, the merging unit includes:

定位子单元，用于确定每个子区域的链接区域；Locating subunits for determining link areas for each subarea;

计算子单元，用于计算所述每个子区域的链接区域的合并权值；a calculation subunit, configured to calculate the combined weight of the link areas of each sub-area;

合并子单元，用于将所述每个子区域与其合并权值满足条件的链接区域进行合并。The merging sub-unit is configured to merge each sub-area with its link area whose merging weight meets a condition.

进一步，所述定位子单元包括：Further, the positioning subunit includes:

提取模块，用于提取子区域与其邻接子区域的特征值；An extraction module, used to extract the feature values of the sub-region and its adjacent sub-regions;

第一计算模块，用于根据提取的特征值，计算所述子区域与其邻接子区域之间的特征相似度；The first calculation module is used to calculate the feature similarity between the sub-region and its adjacent sub-regions according to the extracted feature value;

确定模块，用于将特征相似度大于相似度阈值，且区域面积大于所述子区域的邻接子区域作为所述子区域的链接区域。A determination module, configured to use an adjacent sub-area whose feature similarity is greater than a similarity threshold and whose area is larger than the sub-area as a link area of the sub-area.

进一步，所述计算子单元，具体用于根据所述每个子区域与其链接区域间的特征相似度及语义相似度评估函数，按照网页级别算法，确定所述每个子区域的链接区域的合并权值。Further, the calculation subunit is specifically used to determine the combined weight of the link areas of each sub-area according to the web page level algorithm according to the feature similarity and semantic similarity evaluation function between each sub-area and its link area .

进一步，所述合并子单元，具体用于将所述每个子区域与其合并权值最大的链接区域进行合并。Further, the merging subunit is specifically configured to merge each sub-area with the link area with the largest merging weight.

进一步，所述分割单元包括：Further, the segmentation unit includes:

设定子单元，用于设定最终聚类数目；Set subunits, used to set the final number of clusters;

分割子单元，用于基于所述超图将所述至少两种特征的子分割结果按照所述最终聚类数目进行谱聚类集成，获得所述给定图像的分割结果。The segmentation subunit is configured to perform spectral clustering integration on the sub-segmentation results of the at least two features according to the final number of clusters based on the hypergraph, to obtain the segmentation result of the given image.

进一步，所述设定子单元包括：Further, the setting subunit includes:

第二计算模块，用于计算所述每种特征对应的子分割结果的特征相似度；The second calculation module is used to calculate the feature similarity of the sub-segmentation results corresponding to each feature;

处理模块，用于对所述至少两种特征的子分割结果的特征相似度进行归一化；A processing module, configured to normalize the feature similarity of the sub-segmentation results of the at least two features;

设定模块，用于将归一化后特征相似度的最小值所对应的子分割结果的聚类数目作为所述最终聚类数目。The setting module is used to use the cluster number of the sub-segmentation result corresponding to the minimum value of the normalized feature similarity as the final cluster number.

本发明实施例通过首先获得多种特征的子分割结果，然后将不同特征的子分割结果表示在一个超图中，进而利用聚类集成的方法获得了最终的图像分割结果。该分割方法有效地结合了多种特征对图像的分割，克服了一般分割算法因为基于单一的图像特征而不具备通用性的缺点，同时，通过一个超图模型将多种特征空间下的分割结果进行合并，通过超图上的聚类集成形成对图像的一个最终的划分，避免了因为视觉特征向量和Bag of words中各部分权值和维数的设置不当而使得特征对分割的性能无法起到应有作用的问题，对噪声、异常点、采样点变动不敏感，具有较强的鲁棒性。In the embodiment of the present invention, the sub-segmentation results of various features are first obtained, and then the sub-segmentation results of different features are represented in a hypergraph, and then the final image segmentation result is obtained by using a cluster integration method. This segmentation method effectively combines multiple features for image segmentation, overcomes the shortcomings of general segmentation algorithms that are not universal because they are based on a single image feature, and at the same time, through a hypergraph model, the segmentation results in multiple feature spaces Merge and form a final division of the image through the cluster integration on the hypergraph, avoiding the inability of the feature to segment performance due to the improper setting of the weight and dimension of each part in the visual feature vector and Bag of words It is not sensitive to noise, abnormal points, and sampling point changes, and has strong robustness.

附图说明Description of drawings

为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention. For those skilled in the art, other drawings can also be obtained according to these drawings without any creative effort.

图1是本发明实施例一种图像分割的方法流程图；FIG. 1 is a flowchart of a method for image segmentation according to an embodiment of the present invention;

图2是本发明实施例一种对多个子区域进行合并的方法流程图；Fig. 2 is a flow chart of a method for merging multiple sub-regions according to an embodiment of the present invention;

图3是本发明实施例一种确定某一子区域的链接区域的方法流程图；Fig. 3 is a flow chart of a method for determining a link area of a certain sub-area according to an embodiment of the present invention;

图4是本发明实施例中将三种特征的子分割结果表示为超图的示意图；FIG. 4 is a schematic diagram of representing the sub-segmentation results of three features as a hypergraph in an embodiment of the present invention;

图5是本发明实施例一种基于超图对子分割结果进行谱聚类集成的方法流程图；5 is a flow chart of a method for spectral clustering integration of sub-segmentation results based on a hypergraph according to an embodiment of the present invention;

图6是本发明实施例一种设定最终聚类数目的方法流程图；6 is a flowchart of a method for setting the final number of clusters according to an embodiment of the present invention;

图7是本发明实施例一种图像分割系统的结构示意图；7 is a schematic structural diagram of an image segmentation system according to an embodiment of the present invention;

图8是本发明实施例一种合并单元的结构示意图；FIG. 8 is a schematic structural diagram of a merging unit according to an embodiment of the present invention;

图9是本发明实施例一种分割单元的结构示意图。FIG. 9 is a schematic structural diagram of a segmentation unit according to an embodiment of the present invention.

具体实施方式Detailed ways

为了使本领域技术人员能进一步了解本发明的特征及技术内容，请参阅以下有关本发明的详细说明与附图，附图仅提供参考与说明，并非用来限制本发明。In order for those skilled in the art to further understand the features and technical contents of the present invention, please refer to the following detailed description and accompanying drawings of the present invention. The accompanying drawings are provided for reference and illustration only, and are not intended to limit the present invention.

下面结合附图和实施例，对本发明的技术方案进行描述。The technical solutions of the present invention will be described below in conjunction with the drawings and embodiments.

参见图1，为本发明实施例一种图像分割的方法流程图。Referring to FIG. 1 , it is a flowchart of an image segmentation method according to an embodiment of the present invention.

该方法可以包括：The method can include:

步骤101，对给定图像进行预分割，获得多个子区域。Step 101, pre-segment a given image to obtain multiple sub-regions.

在本发明实施例中，首先对给定图像进行预分割，分割出多个子区域。分割的方法可以参见多种现有技术，例如，利用名称为“Normalized cuts andimage segmentation”，作者为J.Shi,and J.Malik.，发表于IEEE Transactions onPattern Analysis and Machine Intelligence,22(8)(2000)888-905的参考文献提出的方法，即Ncut方法将图像分割成若干个网格状的连接区域。In the embodiment of the present invention, firstly, a given image is pre-segmented, and a plurality of sub-regions are segmented. The method of segmentation can refer to various existing technologies, for example, using the name "Normalized cuts and image segmentation", authored by J.Shi, and J.Malik., published in IEEE Transactions onPattern Analysis and Machine Intelligence, 22(8)( 2000) The method proposed by the reference of 888-905, that is, the Ncut method divides the image into several grid-like connected regions.

步骤102，按照至少两种特征分别对多个子区域进行合并，获得每种特征对应的子分割结果。Step 102, combining multiple sub-regions according to at least two features to obtain a sub-segmentation result corresponding to each feature.

首先，选取至少两种特征，例如颜色、纹理、SIFT（Scale-invariant featuretransform，尺度不变特征转换）等，然后基于不同的特征分别对预分割获得的多个子区域进行合并，获得每种特征对应的子分割结果，例如按照颜色特征对预分割获得的多个子区域进行合并，合并后的结果即为该颜色特征对应的子分割结果。First, select at least two features, such as color, texture, SIFT (Scale-invariant feature transform, scale-invariant feature transform), etc., and then merge the multiple sub-regions obtained by pre-segmentation based on different features to obtain the corresponding The sub-segmentation results of the sub-segmentation results, for example, combining multiple sub-regions obtained by pre-segmentation according to the color feature, and the merged result is the sub-segmentation result corresponding to the color feature.

其中，对于子区域的合并，可以借用网页级别算法或者其他类似于网页级别算法的方式，如基于相似性的随机游走方法，对子区域进行合并，还可以是采用基于语义相似度修正后的网页级别算法对子区域进行合并。具体请参见后续实施例的描述。Among them, for the merging of sub-regions, you can use the page-level algorithm or other methods similar to the page-level algorithm, such as the random walk method based on similarity, to merge the sub-regions, and you can also use the method based on semantic similarity correction. The page-level algorithm merges subregions. For details, please refer to the description of the subsequent embodiments.

步骤103，将该至少两种特征对应的子分割结果表示为超图。Step 103, representing the sub-segmentation results corresponding to the at least two features as a hypergraph.

在获得各特征对应的子分割结果后，即可将所有特征对应的子分割结果表示为一个超图。其中，将预分割获得的各子区域表示为超图的顶点，将各特征对应的子分割结果表示为超图的超边。After obtaining the sub-segmentation results corresponding to each feature, the sub-segmentation results corresponding to all features can be expressed as a hypergraph. Among them, each sub-region obtained by pre-segmentation is represented as a vertex of the hypergraph, and the sub-segmentation results corresponding to each feature are represented as hyperedges of the hypergraph.

具体的转化为超图的过程请参见后续实施例的描述。For the specific process of transforming into a hypergraph, please refer to the description of the subsequent embodiments.

步骤104，基于该超图将至少两种特征的子分割结果进行聚类集成，获得给定图像的分割结果。Step 104, clustering and integrating the sub-segmentation results of at least two features based on the hypergraph to obtain the segmentation result of a given image.

在将各特征的子分割结果表示为超图后，即可依据该超图进行聚类集成，得到该给定图像的最终的分割结果。其中，聚类集成所采用的方法可以是谱聚类，也可以是其它的聚类方法，例如，K-means聚类方法、复杂图方法等，此处不再一一列举。After the sub-segmentation results of each feature are expressed as a hypergraph, clustering and integration can be performed according to the hypergraph to obtain the final segmentation result of the given image. Among them, the method used for clustering integration may be spectral clustering, or other clustering methods, such as K-means clustering method, complex graph method, etc., which will not be listed here.

本发明实施例通过首先获得多种特征的子分割结果，然后将不同特征的子分割结果表示在一个超图中，进而利用聚类集成的方法获得了最终的图像分割结果。该分割方法有效地结合了多种特征对图像的分割，克服了一般分割算法因为基于单一的图像特征而不具备通用性的缺点，同时，通过一个超图模型将多种特征空间下的分割结果进行合并，通过超图上的聚类集成形成对图像的一个最终的划分，避免了因为视觉特征向量和Bag of words中各部分权值和维数的设置不当而使得特征对分割的性能无法起到应有作用的问题，对噪声、异常点、采样点变动不敏感，具有较强的鲁棒性。本发明实施例提供的是一种通用性非常强的分割方法，针对不同类型的图像，能够保证分割质量的稳定性和通用性，而且，该方法是一种无监督的自动分割技术，无需事先输入大量的训练集，在分割中也无需人工参与，如指定聚类的数量，指定区域增长的种子数量等。In the embodiment of the present invention, the sub-segmentation results of various features are first obtained, and then the sub-segmentation results of different features are represented in a hypergraph, and then the final image segmentation result is obtained by using a cluster integration method. This segmentation method effectively combines multiple features for image segmentation, overcomes the shortcomings of general segmentation algorithms that are not universal because they are based on a single image feature, and at the same time, through a hypergraph model, the segmentation results in multiple feature spaces Merge and form a final division of the image through the cluster integration on the hypergraph, avoiding the inability of the feature to segment performance due to the improper setting of the weight and dimension of each part in the visual feature vector and Bag of words It is not sensitive to noise, abnormal points, and sampling point changes, and has strong robustness. The embodiment of the present invention provides a very versatile segmentation method, which can ensure the stability and versatility of the segmentation quality for different types of images. Moreover, the method is an unsupervised automatic segmentation technology without prior Input a large number of training sets, and there is no need for manual participation in the segmentation, such as specifying the number of clusters, specifying the number of seeds for region growth, etc.

在执行步骤102按照至少两种特征分别对多个子区域进行合并，获得每种特征对应的子分割结果时，对于每种特征其合并子区域获得子分割结果的步骤相同，在本发明的另一实施例中，以一种特征为例，如颜色特征，其对多个子区域进行合并的过程，如图2所示，可以包括：When performingstep 102 to merge a plurality of sub-regions according to at least two features to obtain the sub-segmentation results corresponding to each feature, the steps for merging sub-regions to obtain sub-segmentation results for each feature are the same. In another aspect of the present invention In the embodiment, taking a feature as an example, such as a color feature, the process of merging multiple sub-regions, as shown in Figure 2, may include:

步骤201，确定每个子区域的链接区域。Step 201, determine the link area of each sub-area.

在本实施例中，可以根据子区域间的特征相似度和面积来确定每个子区域的链接区域，以确定某一子区域的链接区域为例，其具体过程如图3所示，可以包括：In this embodiment, the link area of each sub-area can be determined according to the feature similarity and area between the sub-areas. Taking the link area of a certain sub-area as an example, the specific process is shown in Figure 3, which can include:

步骤301，提取子区域与其邻接子区域的特征值。Step 301, extracting feature values of the sub-region and its adjacent sub-regions.

在本实施例中所选取的特征为颜色，则在本步骤中即提取子区域与其邻接子区域的颜色特征值，该提取过程可以参见现有技术中的特征值提取方法，此处不再赘述。其中，与某一子区域位置相邻的子区域即为该某一子区域的邻接子区域。In this embodiment, the selected feature is color, then in this step, the color feature values of the sub-region and its adjacent sub-regions are extracted. The extraction process can refer to the feature value extraction method in the prior art, and will not be repeated here. . Wherein, a sub-region adjacent to a certain sub-region is an adjacent sub-region of the certain sub-region.

步骤302，计算子区域与其邻接子区域之间的特征相似度。Step 302, calculating the feature similarity between the sub-region and its adjacent sub-regions.

子区域与其邻接子区域之间的特征相似度可以根据两子区域的特征值进行计算，具体的，可以是两子区域的特征值的欧式距离，例如两子区域的颜色特征值的欧氏距离。The feature similarity between a sub-region and its adjacent sub-regions can be calculated based on the eigenvalues of the two sub-regions. Specifically, it can be the Euclidean distance between the eigenvalues of the two sub-regions, such as the Euclidean distance between the color eigenvalues of the two sub-regions .

步骤303，将特征相似度大于相似度阈值，且区域面积大于该子区域的邻接子区域作为该子区域的链接区域。Instep 303, an adjacent sub-region whose feature similarity is greater than a similarity threshold and whose area is larger than the sub-region is used as a link region of the sub-region.

当子区域与其邻接子区域之间的特征相似度大于设定的相似度阈值时，认为该子区域与其邻接子区域之间具有链接关系，然后比较两子区域的区域面积，大于该子区域面积的邻接子区域即认定为该子区域的链接区域。子区域的链接方向指向其链接区域。When the feature similarity between the sub-region and its adjacent sub-region is greater than the set similarity threshold, it is considered that there is a link relationship between the sub-region and its adjacent sub-region, and then compare the area of the two sub-regions, which is greater than the area of the sub-region The adjacent sub-area of is identified as the link area of the sub-area. The linking direction of a subarea points to its linking area.

其中，相似度阈值可以根据经验设定，也可以针对不同的特征分别设定，例如，颜色特征对应的相似度阈值可以设定为所有子区域的HSV颜色的特征相似度均值；纹理特征对应的相似度阈值可以设定为所有子区域的纹理共生矩阵的特征相似度均值；SIFT特征对应的相似度阈值可以设定为所有子区域的SIFT特征相似度均值。Among them, the similarity threshold can be set according to experience, and can also be set separately for different features. For example, the similarity threshold corresponding to the color feature can be set as the feature similarity average value of the HSV color of all sub-regions; The similarity threshold can be set as the average feature similarity of the texture co-occurrence matrix of all sub-regions; the similarity threshold corresponding to the SIFT feature can be set as the average SIFT feature similarity of all sub-regions.

通过上述步骤301~302即可确定每个子区域的链接区域，在其它实施例中，也可以采用其它方法确定每个子区域的链接区域，例如选用邻接的边界大小等来确定。Through the above steps 301-302, the link area of each sub-area can be determined. In other embodiments, other methods can also be used to determine the link area of each sub-area, for example, the size of the adjacent border is selected to determine.

在确定链接区域后，执行以下步骤202。After the link area is determined, the followingstep 202 is performed.

步骤202，计算每个子区域的链接区域的合并权值。Step 202, calculating the combined weight of the linked areas of each sub-area.

当前图像的分割主要考虑的是图像的底层视觉特征，这些特征只是图像的一种属性，而不具有任何语义意义，与人的语义理解层面上不具备一一对应的关系，也就是语义鸿沟问题，即底层视觉特征相似，并不代表语义内容一致；而语义一致的对象、区域也可能在底层视觉特征上差别很大。The current image segmentation mainly considers the underlying visual features of the image. These features are only an attribute of the image and do not have any semantic meaning. There is no one-to-one correspondence with human semantic understanding, which is the problem of semantic gap. , that is, similar underlying visual features do not mean consistent semantic content; and objects and regions with consistent semantics may also have great differences in underlying visual features.

为了解决视觉底层特征与高层语义之间的语义鸿沟问题，本实施例引入了语义相似度来替代特征相似度，即采用了语义相似度评估函数，将特征相似度输入语义相似度评估函数，该函数有效地模拟了特征相似性与语义相似度的关系，即特征相似不一定语义相似，而语义相似在特征上也可能是不同的。In order to solve the problem of the semantic gap between the visual low-level features and the high-level semantics, this embodiment introduces semantic similarity instead of feature similarity, that is, a semantic similarity evaluation function is used, and the feature similarity is input into the semantic similarity evaluation function. The function effectively simulates the relationship between feature similarity and semantic similarity, that is, similar features are not necessarily semantically similar, and semantic similarity may also be different in terms of features.

在引入了语义相似度后，本实施例采用了类似于网页节点使用的网页级别（Pagerank）算法确定链接区域的合并权值，进而实现区域合并，该方法不仅考虑区域的邻接关系，更是将邻接区域的合并转换为基于语义的合并过程，极大的提高了图像分割的性能。After introducing semantic similarity, this embodiment adopts the Pagerank algorithm similar to that used by webpage nodes to determine the merging weight of link regions, and then realize region merging. This method not only considers the adjacency of regions, but also The merging of adjacent regions is transformed into a semantic-based merging process, which greatly improves the performance of image segmentation.

具体的，可以根据每个子区域与其链接区域间的特征相似度及语义相似度评估函数，按照网页级别算法，确定每个子区域的链接区域的合并权值。Specifically, according to the feature similarity and semantic similarity evaluation function between each sub-area and its link area, and according to the web page level algorithm, the combined weight of the link area of each sub-area can be determined.

首先，基于类似于网页节点使用的Pagerank算法，如果子区域p_i链接指向了子区域p_j，也即子区域p_j为子区域p_i的链接区域，则子区域p_i分配给子区域p_j的语义相似度为：子区域p_i和p_j的语义相似度占p_i和其所有链接区域的语义相似度之和的比率，也即：First, based on the Pagerank algorithm similar to that used by web page nodes, if the sub-area p_i links to the sub-area p_j , that is, the sub-area p_j is the link area of the sub-area p_i , then the sub-area p_i is assigned to the sub-area p The semantic similarity of_j is: the ratio of the semantic similarity of subregions p_i and p_j to the sum of the semantic similarities of p_i and all its linked regions, that is:

${σ σ}_{s the s} (({p p}_{i i},, {p p}_{j j})) / / \underset{k k = = N N (({p p}_{i i}))}{Σ Σ} {σ σ}_{s the s} (({p p}_{k k},, {p p}_{i i}))$

其中，σ_s(p_i,p_j)为子区域p_i和p_j的语义相似度，N(p_i)是一个函数，即求子区域p_i的所有链接区域，子区域p_i的每一个链接区域用p_k表示。Among them, σ_s (p_i ,p_j ) is the semantic similarity between sub-region p_i and p_j , N(p_i ) is a function, that is to find all link regions of sub-region p_i , each sub-region p_i A linked region is denoted by p_k .

本实施例利用子区域之间的视觉特征相似度（如颜色、纹理等特征相似度）来近似子区域之间的语义相似度，即：In this embodiment, the similarity of visual features between sub-regions (such as the similarity of features such as color and texture) is used to approximate the semantic similarity between sub-regions, namely:

${σ σ}_{s the s} (({p p}_{i i},, {p p}_{j j})) &cong; &cong; E E. ((sim sim (({p p}_{i i},, {p p}_{j j}))))$

其中，sim(p_i,p_j)是p_i,p_j两个链接区域的视觉特征相似度（如颜色特征相似度，纹理特征相似度，SIFT特征相似度等）；E是一个语义相似度评估函数，E可以选择正态分布函数或折线函数，当然也可能选用其他的评估函数进行替代。Among them, sim(p_i , p_j ) is the similarity of visual features between p_i and p_j (such as color feature similarity, texture feature similarity, SIFT feature similarity, etc.); E is a semantic similarity Evaluation function, E can choose normal distribution function or broken line function, of course, it is also possible to choose other evaluation functions to replace.

然后，将图像的子区域合并可以看成是由一个节点按照链接节点的语义相似度概率跳转到另一个节点的过程，则采用类似于Pagerank算法中的跳转概率，即可获得每个子区域的链接区域的合并权值P_R，即：Then, merging the sub-regions of the image can be regarded as the process of jumping from one node to another node according to the semantic similarity probability of the link node, and using the jump probability similar to the Pagerank algorithm, each sub-region can be obtained The combined weight P_R of the link area of , namely:

${P P}_{R R} (({p p}_{j j})) = = \frac{((11 - - ϵ ϵ))}{n no} + + ϵ ϵ \times \times \underset{{p p}_{i i} = = N N (({p p}_{j j}))}{Σ Σ} \frac{{P P}_{R R} (({p p}_{i i})) \times \times E E. ((sim sim (({p p}_{i i},, {p p}_{j j}))))}{{Σ Σ}_{{p p}_{k k} = = N N (({p p}_{i i}))} E E. ((sim sim (({p p}_{k k},, {p p}_{i i}))))}$

其中，P_R(p_j)表示区域P_j的合并权值，ε是一个调节常数，n是对图像进行预分割后获得的子区域的数目，例如实验中设定为50，N(p_j)是一个函数，即求子区域p_j的所有链接区域，子区域p_j的每一个链接区域用p_k表示。Among them, P_R (p_j ) represents the combined weight of region P_j , ε is an adjustment constant, n is the number of sub-regions obtained after pre-segmenting the image, for example, it is set to 50 in the experiment, N(p_j ) is a function, that is to find all the link areas of the sub-area p_j , and each link area of the sub-area p_j is denoted by p_k .

在基于上述公式获得每个子区域的链接区域的合并权值P_R后，执行以下步骤203。After the combined weight_PR of the linking areas of each sub-area is obtained based on the above formula, the followingstep 203 is performed.

步骤203，将每个子区域与其合并权值满足条件的链接区域进行合并。Step 203, merging each sub-region with its link region whose merging weight meets the conditions.

在按照上步骤的方法计算获得每个子区域的链接区域的合并权值后，在本实施例中采用了贪心的策略，即将子区域与其合并权值最大的链接区域进行合并，因此加快了合并的速度。在其它实施例中，该合并权值满足的条件可以根据需要或经验设定。After calculating the merging weights of the linking areas of each sub-area according to the method in the previous step, a greedy strategy is adopted in this embodiment, that is, merging the sub-areas with the linking area with the largest merging weight, thus speeding up the process of merging speed. In other embodiments, the conditions to be satisfied by the combination weight can be set according to needs or experience.

在执行上述步骤201~203后，即可获得颜色特征对应的子分割结果，同理，依据上述步骤可以获得纹理特征对应的子分割结果，SIFT特征对应的子分割结果。After performing the above steps 201-203, the sub-segmentation results corresponding to the color features can be obtained. Similarly, the sub-segmentation results corresponding to the texture features and the sub-segmentation results corresponding to the SIFT features can be obtained according to the above steps.

在获得若干特征对应的子分割结果后，即可将这些子分割结果表示为超图，将预分割获得的各子区域表示为超图的顶点，将每种特征对应的子分割结果表示为超图的超边。该构建超图的方法与现有技术类似，可参见名称为“Cluster ensembles-a knowledge reuse framework for combining multiplepartitions”，作者为A.Strehl and J.Ghosh.，发表于The Journal of MachineLearning Research,3(2003)583-617的参考文献。After obtaining the sub-segmentation results corresponding to several features, these sub-segmentation results can be represented as a hypergraph, each sub-region obtained by pre-segmentation is represented as a vertex of the hypergraph, and the sub-segmentation results corresponding to each feature are represented as a hypergraph The hyperedge of the graph. The method for constructing a hypergraph is similar to the prior art, see the title "Cluster ensembles-a knowledge reuse framework for combining multiple partitions", authored by A. Strehl and J. Ghosh., published in The Journal of Machine Learning Research, 3( 2003) References for 583-617.

在另一实施例中，以三种特征对应的子分割结果为例进行说明，该三种特征分别为颜色、纹理和SIFT，基于该三种特征的子分割结果表示为超图的过程如图4所示，可以为：In another embodiment, the sub-segmentation results corresponding to three features are used as an example for illustration. The three features are color, texture and SIFT respectively. The process of expressing the sub-segmentation results based on the three features as a hypergraph is shown in the figure 4, it can be:

用颜色标记向量C，纹理标记向量T和SIFT标记向量S，分别表示利用颜色、纹理和SIFT特征获得的子分割结果。例如C,the label vector C[3,1,2,…]T表示了每个子区域所属的聚类类别，如图4左侧所示，子区域1属于第3个类别，子区域2属于第1个类别等。如果基于同一特征下两个子区域ri和rj的类别标记是相同的，则表明在获得子分割结果的过程中，这两个子区域合并成了一个区域。The color-labeled vector C, the texture-labeled vector T and the SIFT-labeled vector S represent the sub-segmentation results obtained using color, texture and SIFT features, respectively. For example, C, the label vector C[3,1,2,…]T represents the cluster category to which each sub-region belongs. As shown on the left side of Figure 4,sub-region 1 belongs to the third category, andsub-region 2 belongs to the third category. 1 category etc. If the class labels of two sub-regions ri and rj based on the same feature are the same, it indicates that the two sub-regions were merged into one region during the process of obtaining the sub-segmentation results.

对于超图G(V,E),超图的顶点V就代表了预分割获得的子区域,r1,r2,and ri∈V。集合E表示超边（hyperedges）且E={{C(p)},{T(q)},{S(r)}}。每个标记向量都具有一定的聚类数目，例如标记向量C有P个聚类，每个聚类都被表示成C(p)，p=1,2,…,P；标记向量T有q个聚类，每个聚类都被表示成T（q），q=1,2,…,q。将各子区域所属的标记向量的聚类类别表示为矩阵，如图4右侧所示，该矩阵中每个列向量，如C(p),T(q),和S(r)，即表示超图中的一条超边，每个列向量中的1表示对应的行即某个子区域属于这个聚类（超边），0表示对应的行即对应的子区域不属于这个聚类（超边），矩阵{C(p)}中的每行相加都为1。通过上述过程即可将各子分割结果表示成了超图的形式。For the hypergraph G(V,E), the vertex V of the hypergraph represents the sub-regions obtained by pre-segmentation, r1, r2, and ri∈V. The set E represents hyperedges and E={{C(p)},{T(q)},{S(r)}}. Each label vector has a certain number of clusters. For example, the label vector C has P clusters, and each cluster is expressed as C(p), p=1,2,...,P; the label vector T has q clusters, each cluster is represented as T(q), q=1,2,...,q. Express the clustering category of the label vectors to which each sub-region belongs as a matrix, as shown on the right side of Figure 4, each column vector in the matrix, such as C(p), T(q), and S(r), namely Represents a hyperedge in the hypergraph, 1 in each column vector indicates that the corresponding row, that is, a certain subregion belongs to this cluster (hyperedge), and 0 indicates that the corresponding row, that is, the corresponding subregion does not belong to this cluster (superedge). side), each row in the matrix {C(p)} sums to 1. Through the above process, each sub-segmentation result can be expressed in the form of a hypergraph.

在将各子分割结果表示为超图后，即可基于该超图对子分割结果进行聚类集成，获得对图像最终的分割结果。该聚类集成的方法有多种，在本发明的一实施例中，采用谱聚类的方法获得最终的分割结果，如图5所示，该方法可以包括：After each sub-segmentation result is expressed as a hypergraph, the sub-segmentation results can be clustered and integrated based on the hypergraph to obtain the final segmentation result of the image. There are multiple methods for this clustering integration. In one embodiment of the present invention, the method for spectral clustering is used to obtain the final segmentation result. As shown in FIG. 5, the method may include:

步骤501，设定最终聚类数目。Step 501, setting the final number of clusters.

在进行谱聚类集成时，需要先设定聚类数目，该聚类数目可以根据经验或其它现有方法设置，也可以根据某一特征的子分割结果对应的聚类数目设定等。在本实施例中，考虑到在分割中，如果邻接的相似的子区域全都聚类成同一个object，则这个object类内的区域之间的相似度之和应该是最小的，而对于一个好的分割结果，即图像中的所有区域都正确聚类成若干个object，则这个图像的所有聚类的特征相似度之和是最小的。由于每种特征在子分割中分割效果不同，最优的子分割结果应该是能够产生最小相似度之和的子分割结果，而该子分割结果所对应的聚类数目将用于最终聚类集成的聚类数目。如图6所示，其具体过程可以包括：When performing spectral clustering integration, it is necessary to set the number of clusters first. The number of clusters can be set according to experience or other existing methods, or it can be set according to the number of clusters corresponding to the sub-segmentation results of a certain feature, etc. In this embodiment, considering that in segmentation, if adjacent similar subregions are all clustered into the same object, the sum of the similarities between regions within this object class should be the smallest, and for a good The segmentation result of the image, that is, all regions in the image are correctly clustered into several objects, then the sum of the feature similarities of all clusters of this image is the smallest. Since each feature has different segmentation effects in the sub-segmentation, the optimal sub-segmentation result should be the sub-segmentation result that can produce the minimum sum of similarities, and the number of clusters corresponding to the sub-segmentation result will be used for the final clustering integration number of clusters. As shown in Figure 6, the specific process may include:

步骤601，针对至少两种特征对应的子分割结果分别计算特征相似度。In step 601, feature similarities are calculated for sub-segmentation results corresponding to at least two features.

以颜色、纹理和SIFT三种特征为例，分别计算各特征下获得的子分割结果的特征相似度，也即计算合并区域后区域之间在某一特征下的特征相似度，该具体计算方法可以与前述实施例中的步骤301、302类似，此处不再赘述。Taking the three features of color, texture and SIFT as examples, calculate the feature similarity of the sub-segmentation results obtained under each feature, that is, calculate the feature similarity between the merged regions under a certain feature. The specific calculation method It may be similar to steps 301 and 302 in the foregoing embodiments, and details are not repeated here.

步骤602，对计算获得的特征相似度进行归一化。Step 602, normalize the calculated feature similarity.

由于每个子分割结果使用了不同的特征，因此所计算出来的特征相似度不在同一量纲中，可以通过高斯函数对计算出的特征相似的进行归一化。Since each sub-segmentation result uses different features, the calculated feature similarity is not in the same dimension, and the calculated feature similarity can be normalized by a Gaussian function.

步骤603，将归一化后特征相似度的最小值所对应的子分割结果的聚类数目作为最终聚类数目。Step 603, taking the cluster number of the sub-segmentation result corresponding to the minimum value of the normalized feature similarity as the final cluster number.

在归一化后，即可将特征相似度最小的子分割结果所对应的聚类数目作为最终进行谱聚类的聚类数目，例如，颜色特征下对应的子分割结果的特征相似度最小，则该子分割结果的聚类数目p，即可作为最终的数目。After normalization, the number of clusters corresponding to the sub-segmentation result with the smallest feature similarity can be used as the number of clusters for the final spectral clustering. For example, the feature similarity of the sub-segmentation result corresponding to the color feature is the smallest, Then the cluster number p of the sub-segmentation result can be used as the final number.

在确定谱聚类的聚类数目后，即可执行以下步骤502。After the number of clusters in spectral clustering is determined, the following step 502 can be performed.

步骤502，基于超图将各子分割结果按照最终聚类数目进行谱聚类集成，获得给定图像的分割结果。Step 502 , based on the hypergraph, perform spectral clustering integration of each sub-segmentation result according to the final cluster number, to obtain the segmentation result of a given image.

基于谱聚类的原理，当两个区域被划分到一个类时，表明这两个区域在多种特征下都进行了合并，因此集成的结果是这两个区域最终将被合并成一个区域；相反，如果存在两个区域在少数特征的子分割过程中进行了合并，而在大多数特征下都没有合并，则表明，这个两个区域是存在差异的，不应该进行合并，则集成的结果是这两个区域不会被合并。例如，两个绿色区域分别代表了草地和灌木，在基于颜色特征的情况下，进行了合并，而在纹理和SIFT特征空间下都没有进行合并，则集成的结果是也没有合并。具体的根据超图模型及设定的聚类数目进行谱聚类的过程与现有技术类似，可参见名称为“On spectral clustering:analysis and an algorithm”，作者为NG A Y，JORDAN M I,WEISS Y.，发表于Advanced in Neural Information ProcessingSystem,2002,2:849-856的文献，此处不再赘述。Based on the principle of spectral clustering, when two regions are classified into one class, it indicates that the two regions have been merged under various features, so the result of the integration is that the two regions will eventually be merged into one region; On the contrary, if there are two regions that are merged in the sub-segmentation process of a few features, but not merged under most features, it indicates that the two regions are different and should not be merged. The result of the integration Yes the two regions will not be merged. For example, two green areas represent grassland and shrubs respectively, and are merged based on color features, but not merged under texture and SIFT feature space, and the result of integration is that there is no merge. The specific process of spectral clustering based on the hypergraph model and the set number of clusters is similar to the existing technology, see the title "On spectral clustering: analysis and an algorithm", authored by NG A Y, JORDAN M I, WEISS Y., a document published in Advanced in Neural Information Processing System, 2002, 2:849-856, will not be repeated here.

经过上述过程即可形成最终的图像分割结果。After the above process, the final image segmentation result can be formed.

本发明实施例通过首先分别从全局特征中选取颜色、纹理特征，从局部尺度不变特征中选择SIFT特征，分别基于颜色、纹理、SIFT特征对图像进行分割，然后通过一个超图模型将多种特征空间下的分割结果进行合并，通过超图上的谱聚类算法形成了对图像的一个最终的划分。该分割方法有效地结合了多种特征对图像进行分割；而且通过将图像的子区域合并可以看成是由一个节点按照链接节点的语义相似度概率跳转到另一个节点的过程，不仅考虑了区域的邻接关系，而且将邻接区域的合并转换为基于语义的合并过程，极大的提高了分割的性能；通过设置合并的链接权值，采用贪心的策略，改善了区域合并的速度；进一步地，通过选取谱聚类的聚类数目，获得了更好的分割效果。In the embodiment of the present invention, the color and texture features are respectively selected from the global features, and the SIFT features are selected from the local scale-invariant features, and the image is segmented based on the color, texture, and SIFT features respectively, and then various The segmentation results in the feature space are combined, and a final division of the image is formed through the spectral clustering algorithm on the hypergraph. This segmentation method effectively combines a variety of features to segment the image; and by merging the sub-regions of the image, it can be regarded as the process of jumping from one node to another node according to the semantic similarity probability of the link node, not only considering The adjacency relationship of the region, and the merging of adjacent regions is converted into a semantic-based merging process, which greatly improves the performance of the segmentation; by setting the link weight of the merger and adopting a greedy strategy, the speed of region merging is improved; further , by selecting the number of clusters of spectral clustering, a better segmentation effect is obtained.

以上是对本发明方法实施例的描述，下面对实现上述方法的系统进行介绍。The above is the description of the method embodiment of the present invention, and the system for realizing the above method will be introduced below.

参见图7，为本发明实施例一种图像分割系统的结构示意图。Referring to FIG. 7 , it is a schematic structural diagram of an image segmentation system according to an embodiment of the present invention.

该系统可以包括：The system can include:

预分割单元701，用于对给定图像进行预分割，获得多个子区域；Apre-segmentation unit 701, configured to pre-segment a given image to obtain multiple sub-regions;

合并单元702，用于按照至少两种特征分别对所述多个子区域进行合并，获得每种特征对应的子分割结果；A mergingunit 702, configured to merge the plurality of sub-regions according to at least two features, to obtain a sub-segmentation result corresponding to each feature;

转换单元703，用于将至少两种特征对应的子分割结果表示为超图；Aconversion unit 703, configured to represent the sub-segmentation results corresponding to at least two features as a hypergraph;

分割单元704，用于基于超图将至少两种特征的子分割结果进行聚类集成，获得给定图像的分割结果。Thesegmentation unit 704 is configured to cluster and integrate the sub-segmentation results of at least two features based on the hypergraph to obtain a segmentation result of a given image.

首先，预分割单元701对给定图像进行预分割，分割出多个子区域。分割的方法可以采用Ncut方法将图像分割成若干个网格状的连接区域，然后，合并单元702基于不同的特征分别对预分割获得的多个子区域进行合并，获得每种特征对应的子分割结果，其中，对于子区域的合并，可以借用网页级别算法或者其他类似于网页级别算法的方式，如基于相似性的随机游走方法，对子区域进行合并。在合并单元702获得多个子分割结果后，转换单元703即可将所有特征对应的子分割结果表示为一个超图。其中，将预分割获得的各子区域表示为超图的顶点，将各特征对应的子分割结果表示为超图的超边。最后由分割单元704依据该超图进行聚类集成，得到该给定图像的最终的分割结果。其中，聚类集成所采用的方法可以是谱聚类，也可以是其它的聚类方法。First, thepre-segmentation unit 701 performs pre-segmentation on a given image to segment a plurality of sub-regions. The segmentation method can use the Ncut method to divide the image into several grid-like connected areas, and then, the mergingunit 702 combines the multiple sub-areas obtained by pre-segmentation based on different features to obtain the sub-segmentation results corresponding to each feature , wherein, for the merging of sub-regions, the sub-regions can be merged by borrowing the page-level algorithm or other methods similar to the page-level algorithm, such as a similarity-based random walk method. After themerging unit 702 obtains multiple sub-segmentation results, theconversion unit 703 can represent the sub-segmentation results corresponding to all features as a hypergraph. Among them, each sub-region obtained by pre-segmentation is represented as a vertex of the hypergraph, and the sub-segmentation results corresponding to each feature are represented as hyperedges of the hypergraph. Finally, thesegmentation unit 704 performs cluster integration according to the hypergraph to obtain the final segmentation result of the given image. Wherein, the method adopted for the cluster integration may be spectral clustering or other clustering methods.

本发明实施例中，该系统实现的分割方法有效地结合了多种特征对图像的分割，通过一个超图模型将多种特征空间下的分割结果进行合并，并通过超图上的聚类集成形成对图像的一个最终的划分，使得特征对分割的性能起到了应有的作用。In the embodiment of the present invention, the segmentation method implemented by the system effectively combines the segmentation of multiple features on the image, combines the segmentation results in multiple feature spaces through a hypergraph model, and integrates the results through clustering on the hypergraph. A final division of the image is formed, so that the features play a due role in the performance of the segmentation.

参见图8，为本发明实施例一种合并单元的结构示意图。Referring to FIG. 8 , it is a schematic structural diagram of a merging unit according to an embodiment of the present invention.

在本发明实施例中，该图像分割系统的合并单元可以包括：In an embodiment of the present invention, the merging unit of the image segmentation system may include:

定位子单元801，用于确定每个子区域的链接区域；Apositioning subunit 801, configured to determine the link area of each sub-area;

计算子单元802，用于计算每个子区域的链接区域的合并权值；A calculatingsubunit 802, configured to calculate the combined weight of the link areas of each sub-area;

合并子单元803，用于将每个子区域与其合并权值满足条件的链接区域进行合并。The mergingsubunit 803 is configured to merge each sub-area with its link area whose merging weight meets the conditions.

其中，该定位子单元801还可以进一步包括：Wherein, thepositioning subunit 801 may further include:

提取模块8011，用于提取子区域与其邻接子区域的特征值；Extraction module 8011, used to extract the feature values of the sub-region and its adjacent sub-regions;

第一计算模块8012，用于根据提取的特征值，计算子区域与其邻接子区域之间的特征相似度；Thefirst calculation module 8012 is used to calculate the feature similarity between the sub-region and its adjacent sub-regions according to the extracted feature value;

确定模块8013，用于将特征相似度大于相似度阈值，且区域面积大于子区域的邻接子区域作为该子区域的链接区域。The determiningmodule 8013 is configured to use the adjacent sub-region whose feature similarity is greater than the similarity threshold and whose area is larger than the sub-region as the link region of the sub-region.

计算子单元802，具体可以用于根据每个子区域与其链接区域间的特征相似度及语义相似度评估函数，按照网页级别算法，确定每个子区域的链接区域的合并权值。Thecalculation subunit 802 can specifically be used to determine the combined weight of the link areas of each sub-area according to the web page level algorithm according to the feature similarity and semantic similarity evaluation function between each sub-area and its link areas.

合并子单元803，具体可以用于将每个子区域与其合并权值最大的链接区域进行合并。The mergingsubunit 803 can be specifically configured to merge each sub-region with the link region with the largest merging weight.

该系统的合并单元通过将图像的子区域合并可以看成是由一个节点按照链接节点的语义相似度概率跳转到另一个节点的过程，不仅考虑了区域的邻接关系，而且将邻接区域的合并转换为基于语义的合并过程，极大的提高了分割的性能；并且通过设置合并的链接权值，采用贪心的策略，改善了区域合并的速度。The merging unit of the system can be regarded as the process of jumping from one node to another node according to the semantic similarity probability of the link node by merging the sub-regions of the image, not only considering the adjacency of the region, but also merging the adjacent regions Converting to a semantic-based merging process greatly improves the performance of segmentation; and by setting the merging link weights and adopting a greedy strategy, the speed of region merging is improved.

参见图9，为本发明实施例一种分割单元的结构示意图。Referring to FIG. 9 , it is a schematic structural diagram of a segmentation unit according to an embodiment of the present invention.

在本发明实施例中，该图像分割系统的分割单元可以包括：In an embodiment of the present invention, the segmentation unit of the image segmentation system may include:

设定子单元901，用于设定最终聚类数目；A settingsubunit 901 is used to set the final number of clusters;

分割子单元902，用于基于超图将至少两种特征的子分割结果按照最终聚类数目进行谱聚类集成，获得所述给定图像的分割结果。Thesegmentation subunit 902 is configured to perform spectral clustering integration of the sub-segmentation results of at least two features according to the final cluster number based on the hypergraph, and obtain the segmentation result of the given image.

其中，该设定子单元901还可以进一步包括：Wherein, the settingsubunit 901 may further include:

第二计算模块9011，用于计算每种特征对应的子分割结果的特征相似度；Thesecond calculation module 9011 is used to calculate the feature similarity of the sub-segmentation results corresponding to each feature;

处理模块9012，用于对至少两种特征的子分割结果的特征相似度进行归一化；Aprocessing module 9012, configured to normalize the feature similarity of the sub-segmentation results of at least two features;

设定模块9013，用于将归一化后特征相似度的最小值所对应的子分割结果的聚类数目作为最终聚类数目。Thesetting module 9013 is configured to use the cluster number of the sub-segmentation result corresponding to the minimum value of the normalized feature similarity as the final cluster number.

该分割单元通过选取谱聚类的聚类数目，并通过超图上的谱聚类算法形成了对图像的一个最终的划分，获得了更好的分割效果。The segmentation unit forms a final division of the image through the spectral clustering algorithm on the hypergraph by selecting the cluster number of the spectral clustering, and obtains a better segmentation effect.

以上系统中的各单元的具体实现过程请参照前述方法实施例的相应描述，此处不再赘述。For the specific implementation process of each unit in the above system, please refer to the corresponding description of the foregoing method embodiments, and details are not repeated here.

以上所述的本发明实施方式，并不构成对本发明保护范围的限定。任何在本发明的精神和原则之内所作的修改、等同替换和改进等，均应包含在本发明的权利要求保护范围之内。The embodiments of the present invention described above are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the claims of the present invention.