CN103295013A - Pared area based single-image shadow detection method - Google Patents

Abstract

Translated fromChinese

本发明公开了一种基于成对区域的单幅图像阴影检测方法，包括以下步骤：对已知的自然场景阴影图像利用中值滤波方法进行预处理，获取预处理后阴影图像；利用聚类方法对预处理后阴影图像进行初始分割，获取分割后阴影图像区域；对分割后阴影图像区域进行合并，得到最终分割结果；对分割后的各个区域提取特征值，并获取特征向量；对特征向量进行归一化处理，建立阴影区域分类模型；对未知的自然场景阴影图像样本重复操作，提取未知图像特征值并作为支持向量机的输入量，根据阴影区域分类模型对图像区域进行判决，将不同区域分为阴影区域或非阴影区域。本发明降低了计算量，提高了检测效率，能够更加准确高效地检测复杂自然纹理图像的阴影区域。

The invention discloses a single image shadow detection method based on paired regions, comprising the following steps: performing preprocessing on a known natural scene shadow image using a median filtering method to obtain a preprocessed shadow image; using a clustering method Initially segment the preprocessed shadow image to obtain the segmented shadow image area; merge the segmented shadow image area to obtain the final segmentation result; extract feature values from each segmented area and obtain the feature vector; Normalize processing to establish a shadow area classification model; repeat the operation on unknown natural scene shadow image samples, extract unknown image feature values and use them as the input of the support vector machine, judge the image area according to the shadow area classification model, and classify different areas Divided into shaded or non-shaded areas. The invention reduces the calculation amount, improves the detection efficiency, and can detect the shadow area of complex natural texture images more accurately and efficiently.

Description

Translated fromChinese

一种基于成对区域的单幅图像阴影检测方法A Single Image Shadow Detection Method Based on Paired Regions

技术领域technical field

本发明涉及计算机图像处理领域，特别涉及一种基于成对区域的单幅图像阴影检测方法。The invention relates to the field of computer image processing, in particular to a single image shadow detection method based on paired regions.

背景技术Background technique

阴影处理是计算机视觉领域的重要研究内容。作为图像退化的一种特殊形式，阴影的存在使得影像中的目标信息有所损失或受到干扰，在后续的图像处理过程中直接影响到相应区域的边缘提取、目标识别，以及影像匹配等算法的成功率，对航空图像处理、医学成像和视频监控等领域造成很大影响。因此，必须首先对图像中的阴影进行检测和分析，并根据需要消除或减弱阴影的影响。Shadow processing is an important research content in the field of computer vision. As a special form of image degradation, the existence of shadows causes the target information in the image to be lost or disturbed, which directly affects the edge extraction, target recognition, and image matching algorithms of the corresponding area in the subsequent image processing process. The success rate has a great impact on the fields of aerial image processing, medical imaging and video surveillance. Therefore, the shadow in the image must be detected and analyzed first, and the influence of the shadow should be eliminated or weakened as required.

阴影检测是阴影去除的前提和基础，也是该领域的重要研究内容，阴影去除的效果在很大程度上取决于阴影检测效果的好坏。静态图像由于所包含的信息量较少，其阴影检测一直是该领域的难点和主要研究方向。Shadow detection is the premise and foundation of shadow removal, and it is also an important research content in this field. The effect of shadow removal depends to a large extent on the quality of shadow detection. Due to the small amount of information contained in static images, shadow detection has always been a difficult point and the main research direction in this field.

目前，静态图像的阴影检测方法大致可分为：基于模型的方法和基于特征的方法两大类。其中，基于模型的方法是指利用场景、运动目标、光照条件等方面的先验信息，建立阴影模型，例如：Retinex模型及其相关改进方法，以及利用光照条件建立的阴影模型等。基于特征的方法主要是利用阴影的亮度、梯度、色彩、纹理等信息来标识阴影区域，例如：有的科研人员根据静态室内图像阴影的半阴影区域的特征，以及室内环境色调的特点来检测阴影区域的边缘；也有一些科研人员利用图像的色彩特征和纹理属性，采用条件随机场(CRF)标识阴影区域等。At present, shadow detection methods for static images can be roughly divided into two categories: model-based methods and feature-based methods. Among them, the model-based method refers to the use of prior information on scenes, moving objects, and lighting conditions to establish shadow models, such as the Retinex model and its related improved methods, and shadow models established using lighting conditions. Feature-based methods mainly use shadow brightness, gradient, color, texture and other information to identify shadow areas. For example, some researchers detect shadows based on the characteristics of the semi-shadow area of static indoor image shadows and the characteristics of the indoor environment tone. The edge of the area; some researchers also use the color characteristics and texture properties of the image to identify the shadow area by using conditional random field (CRF).

发明人在实现本发明的过程中，发现现有技术中至少存下以下缺点和不足：In the process of realizing the present invention, the inventor finds that there are at least the following shortcomings and deficiencies in the prior art:

基于模型的方法：通常具有比较严密的理论推导，但同时也有较大的局限性，尤其在背景复杂、光照条件较差的条件下，模型的复杂度以及建模时间都会迅速增加，难以满足实际工程的需要；基于特征的方法：对不同的场景以及光照条件具有较强的鲁棒性，但算法的普适性尚有待提高。Model-based methods: usually have relatively rigorous theoretical derivation, but at the same time also have relatively large limitations, especially under the conditions of complex background and poor lighting conditions, the complexity of the model and the modeling time will increase rapidly, which is difficult to meet the actual requirements. Engineering needs; feature-based method: it has strong robustness to different scenes and lighting conditions, but the universality of the algorithm needs to be improved.

发明内容Contents of the invention

本发明提供了一种基于成对区域的单幅图像阴影检测方法，本方法通过对原始自然场景阴影图像进行分割，对分割后的各个区域提取特征值，根据成对分割区域的相关特性，构建支持向量机函数对其进行分类，实现阴影与非阴影区域的自动检测，详见下文描述：The invention provides a single image shadow detection method based on paired regions. This method segments the original natural scene shadow image, extracts feature values from each segmented region, and constructs The support vector machine function classifies it to realize the automatic detection of shadow and non-shadow areas. See the description below for details:

一种基于成对区域的单幅图像阴影检测方法，所述方法包括以下步骤：A single image shadow detection method based on paired regions, said method comprising the following steps:

（1）对已知的自然场景阴影图像利用中值滤波方法进行预处理，获取预处理后阴影图像；(1) Preprocess the known natural scene shadow images using the median filter method to obtain preprocessed shadow images;

（2）利用聚类方法对所述预处理后阴影图像进行初始分割，获取分割后阴影图像区域；(2) Using a clustering method to initially segment the preprocessed shadow image, and obtain the segmented shadow image area;

（3）对所述分割后阴影图像区域进行合并，得到最终分割结果；(3) Merging the segmented shadow image regions to obtain a final segmentation result;

（4）对分割后的各个区域提取特征值，并获取特征向量；(4) Extract eigenvalues for each segmented region and obtain eigenvectors;

（5）对特征向量进行归一化处理，建立阴影区域分类模型；(5) Normalize the eigenvectors to establish a shadow area classification model;

（6）对未知的自然场景阴影图像样本重复步骤（1）～（4），提取未知图像特征值并作为支持向量机的输入量，根据所述阴影区域分类模型对图像区域进行判决，将不同区域分为阴影区域或非阴影区域。(6) Repeat steps (1) to (4) for the shadow image sample of the unknown natural scene, extract the feature value of the unknown image and use it as the input of the support vector machine, and judge the image area according to the shadow area classification model. Areas are classified as shaded or unshaded.

所述利用聚类方法对所述预处理后阴影图像进行初始分割，获取分割后阴影图像区域的步骤具体包括：The step of initially segmenting the preprocessed shadow image by using a clustering method, and obtaining the segmented shadow image region specifically includes:

1）将所述预处理后阴影图像中所有像素划分为K个组；1) Divide all pixels in the preprocessed shadow image into K groups;

2）选择图像中所有像素颜色值的K个等分点作为初始聚类中心，计算N-K个像素点与初始聚类中心的颜色欧氏距离，将N-K个像素点分别分配给欧式距离最小的聚类，形成新聚类；2) Select K equally divided points of all pixel color values in the image as the initial clustering center, calculate the color Euclidean distance between N-K pixel points and the initial clustering center, and assign N-K pixel points to the cluster with the smallest Euclidean distance class to form a new cluster;

3）计算每个所述新聚类的聚类中心，不断重复步骤2）直到每个聚类中心不再变化。3) Calculate the cluster centers of each of the new clusters, and repeat step 2) until each cluster center does not change.

所述对所述分割后阴影图像区域进行合并，得到最终分割结果的步骤具体包括：The step of merging the segmented shadow image regions to obtain the final segmentation result specifically includes:

1）比较分割后各相邻阴影图像区域的特征信息，当两个相互独立的区域满足颜色距离阈值、边缘距离阈值，并且区域连接处没有边缘时，对这两个区域进行合并；1) Compare the feature information of each adjacent shadow image region after segmentation, and merge the two regions when two mutually independent regions meet the color distance threshold and edge distance threshold, and there is no edge at the region connection;

2）在某一区域合并前后，图像区域的颜色散度变化大于散度阈值时，自动停止区域合并过程，得到所述最终分割结果。2) Before and after a region is merged, when the change of the color divergence of the image region is greater than the divergence threshold, the region merge process is automatically stopped, and the final segmentation result is obtained.

所述对分割后的各个区域提取特征值，并获取特征向量的步骤具体包括：The step of extracting feature values from each segmented region and obtaining feature vectors specifically includes:

比较分割后各个区域组成的成对区域的4种特征值，即颜色直方图与纹理直方图、亮度比值、色差角度以及区域在图像中的距离，通过4种特征值得到8个特征向量；其中，颜色直方图与纹理直方图作为2个特征向量；R、G、B三个通道的平均亮度比值，作为3个特征向量：Compare the 4 kinds of eigenvalues of the paired regions composed of each region after segmentation, namely the color histogram and texture histogram, the brightness ratio, the color difference angle and the distance of the region in the image, and obtain 8 eigenvectors through the 4 kinds of eigenvalues; , the color histogram and texture histogram are used as two feature vectors; the average brightness ratio of the three channels of R, G, and B is used as three feature vectors:

${\mathrm{<span><span>P</span>P</span>}}_{\mathrm{<span><span>R</span>R</span>}}<span><span>=</span>=</span>\frac{{\mathrm{<span><span>R</span>R</span>}}_{\mathrm{<span><span>avg</span>avg</span>}\mathrm{<span><span>1</span>1</span>}}}{{\mathrm{<span><span>R</span><figure-callout\; label="R\; avg"\; filenames="HDA00003181746200022.png"\; state="\{\{state\}\}">R</figure-callout></span><figure-callout\; label="R\; avg"\; filenames="HDA00003181746200022.png"\; state="\{\{state\}\}"></figure-callout>}}_{\mathrm{<span><span></span></span>}}}<span><span>,</span>,</span>$${\mathrm{<span><span>P</span>P</span>}}_{\mathrm{<span><span>G</span>G</span>}}<span><span>=</span>=</span>\frac{{\mathrm{<span><span>G</span>G</span>}}_{\mathrm{<span><span>avg</span>avg</span>}\mathrm{<span><span>1</span>1</span>}}}{{\mathrm{<span><span>G</span><figure-callout\; label="G\; avg"\; filenames="HDA00003181746200022.png"\; state="\{\{state\}\}">G</figure-callout></span><figure-callout\; label="G\; avg"\; filenames="HDA00003181746200022.png"\; state="\{\{state\}\}"></figure-callout>}}_{\mathrm{<span><span></span></span>}}}<span><span>,</span>,</span>$${\mathrm{<span><span>P</span>P</span>}}_{\mathrm{<span><span>B</span>B</span>}}<span><span>=</span>=</span>\frac{{\mathrm{<span><span>B</span>B</span>}}_{\mathrm{<span><span>avg</span>avg</span>}\mathrm{<span><span>1</span>1</span>}}}{{\mathrm{<span><span>B</span><figure-callout\; label="B\; avg"\; filenames="HDA00003181746200022.png"\; state="\{\{state\}\}">B</figure-callout></span><figure-callout\; label="B\; avg"\; filenames="HDA00003181746200022.png"\; state="\{\{state\}\}"></figure-callout>}}_{\mathrm{<span><span></span></span>}}}$

其中，R_avg1代表成对区域中第1个区域的红色通道平均值，R_avg2代表成对区域中第2个区域的红色通道平均值，G_avg1代表成对区域中第1个区域的绿色通道平均值，G_avg2代表成对区域中第2个区域的绿色通道平均值，B_avg1代表成对区域中第1个区域的蓝色通道平均值，B_avg2代表成对区域中第2个区域的蓝色通道平均值；Among them, R_avg1 represents the average value of the red channel of the first region in the paired region, R_avg2 represents the average value of the red channel of the second region in the paired region, and G_avg1 represents the green channel of the first region in the paired region Average value, G_avg2 represents the average value of the green channel of the second region in the paired region, B_avg1 represents the average value of the blue channel of the first region in the paired region, B_avg2 represents the average value of the second region in the paired region blue channel average;

采用色差角度P_R/P_G和P_R/P_B作为2个特征向量，由区域面积的几何平均数除以区域中心的欧几里德距离得到1个特征向量，即：Using the color difference angles P_R /_PG and P_R /P_B as two eigenvectors, one eigenvector is obtained by dividing the geometric mean of the area area by the Euclidean distance of the area center, namely:

$\sqrt{{\mathrm{<span><span>s</span>the\; s</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>\&times;</span>\&times;</span>{\mathrm{<span><span>s</span>the\; s</span>}}_{\mathrm{<span><span>j</span>j</span>}}}<span><span>/</span>/</span>\mathrm{<span><span>d</span>d</span>}$

其中，s_i和s_j分别是第i和第j区域的面积，d是第i和第j区域中心的欧几里德距离。Among them, s_i and s_j are the areas of the i-th and j-th regions, respectively, and d is the Euclidean distance between the centers of the i-th and j-th regions.

所述对特征向量进行归一化处理，建立阴影区域分类模型的步骤具体包括：The step of performing normalization processing on the feature vectors and establishing a shadow area classification model specifically includes:

选取一部分经过中值滤波处理的阴影图像作为训练集，对训练集中的所有样本，手动标注阴影区域、成对区域的关系，计算特征向量，将归一化后的特征向量作为SVM的输入量，采用高斯径向基核函数作为支持向量机核函数，建立SVM的最优分类模型。Select a part of the shadow image processed by the median filter as the training set, manually mark the relationship between the shadow area and the paired area for all samples in the training set, calculate the feature vector, and use the normalized feature vector as the input of the SVM, Gaussian radial basis kernel function is used as the kernel function of support vector machine to establish the optimal classification model of SVM.

本发明提供的技术方案的有益效果是：The beneficial effects of the technical solution provided by the invention are:

（1）提出了对初始分割后的纹理区域重新进行合并，以提高图像中各区域的完整性，并通过减少过分割，降低计算量，提高了检测效率。(1) It is proposed to re-merge the texture regions after the initial segmentation to improve the integrity of each region in the image, and to improve the detection efficiency by reducing over-segmentation and reducing the amount of calculation.

（2）将分割后的所有区域与其他各个区域配对，提取成对区域的亮度关系和纹理信息等相关特性，并构建支持向量机函数对其进行分类，利用结构风险最小化原则构建分类模型，比传统的基于阴影特征的方法检测结果更加准确。(2) Pair all the segmented regions with other regions, extract the brightness relationship and texture information and other related characteristics of the paired regions, and construct a support vector machine function to classify them, and construct a classification model using the principle of structural risk minimization, The detection result is more accurate than the traditional method based on shadow features.

总之，本发明将图像分割相关理论用于自然图像的阴影检测，通过与支持向量机相结合，能更加准确高效地检测复杂自然纹理图像的阴影区域。In a word, the present invention applies the relevant theory of image segmentation to the shadow detection of natural images, and can detect shadow regions of complex natural texture images more accurately and efficiently by combining with support vector machines.

附图说明Description of drawings

图1为5幅原始自然场景阴影图像；Figure 1 is five original natural scene shadow images;

图2为利用传统阴影检测算法的阴影检测效果图；Figure 2 is a shadow detection effect diagram using a traditional shadow detection algorithm;

图3为用本发明算法的阴影检测效果图；Fig. 3 is the shadow detection effect diagram with algorithm of the present invention;

图4为SVM对分类模型寻找最优参数，进行分类测试的效果图；Figure 4 is an effect diagram of SVM searching for optimal parameters for classification models and performing classification tests;

图5为本方法的流程图。Figure 5 is a flowchart of the method.

具体实施方式Detailed ways

为使本发明的目的、技术方案和优点更加清楚，下面将结合附图对本发明实施方式作进一步地详细描述。In order to make the object, technical solution and advantages of the present invention clearer, the implementation manner of the present invention will be further described in detail below in conjunction with the accompanying drawings.

本发明针对传统算法中采用的均值漂移算法对局部细节、噪声比较敏感，容易对具有复杂纹理信息的自然场景图像产生过分割的缺点，利用聚类的方法对图像各分割区域进行合并，以提高算法的检测效果和效率，参见图5，详见下文描述：The invention aims at the disadvantage that the mean shift algorithm adopted in the traditional algorithm is sensitive to local details and noise, and is easy to over-segment natural scene images with complex texture information, and uses a clustering method to merge each segmented area of the image to improve For the detection effect and efficiency of the algorithm, see Figure 5, and see the description below for details:

101：对已知的自然场景阴影图像利用中值滤波方法进行预处理，获取预处理后阴影图像；101: Preprocessing the known shadow images of natural scenes using the median filter method to obtain the preprocessed shadow images;

该步骤采用经典中值滤波算法对自然场景阴影图像进行预处理，中值滤波将每一像素点的灰度值设置为该点某邻域窗口内的所有像素点灰度值的中值，从而消除孤立的噪声点。本实验采用的邻域窗口的大小为3(像素)×3(像素)，实现了对阴影图像的细节区域的平滑，以减少畸变点的影响。In this step, the classic median filtering algorithm is used to preprocess the shadow image of the natural scene. The median filtering sets the gray value of each pixel to the median of all the gray values of the pixels in a certain neighborhood window of the point, so that Remove isolated noise points. The size of the neighborhood window used in this experiment is 3 (pixels) × 3 (pixels), which achieves smoothing of the detailed area of the shadow image to reduce the influence of distortion points.

102：利用聚类方法对预处理后阴影图像进行初始分割，获取分割后阴影图像区域；102: Using a clustering method to initially segment the preprocessed shadow image, and obtain the segmented shadow image area;

在R、G、B颜色空间，利用K-means聚类方法对预处理后阴影图像进行分割，得到初始分割的结果。K-means聚类方法将事先输入的n个数据对象划分为k个聚类以使所获得的聚类满足：同一聚类中的对象相似度较高；而不同聚类中的对象相似度较小。具体作法是：In the R, G, and B color spaces, the K-means clustering method is used to segment the preprocessed shadow image to obtain the initial segmentation results. The K-means clustering method divides the input n data objects into k clusters so that the obtained clusters satisfy: the similarity of objects in the same cluster is high; while the similarity of objects in different clusters is relatively low. Small. The specific method is:

1）将预处理后阴影图像中所有像素划分为K个组；1) Divide all pixels in the preprocessed shadow image into K groups;

其中，K<N，N为图像的像素总数，本发明以K=24为例进行说明。Wherein, K<N, N is the total number of pixels of the image, and the present invention takes K=24 as an example for illustration.

2）选择图像中所有像素颜色值的K(即24)个等分点作为初始聚类中心，计算N-K个像素点与初始聚类中心的颜色欧氏距离，将N-K个像素点分别分配给欧式距离最小的聚类，形成新聚类；2) Select K (that is, 24) equally divided points of all pixel color values in the image as the initial clustering center, calculate the color Euclidean distance between N-K pixel points and the initial clustering center, and assign N-K pixel points to the Euclidean The cluster with the smallest distance forms a new cluster;

例如：N的取值为200，将24个等分点分别作为初始聚类中心，计算176个像素点与初始聚类中心的颜色欧式距离，即对每个像素点都计算24个颜色欧式距离，从24个颜色欧氏距离中查找最小欧氏距离，将A像素点归到最小欧氏距离所对应的初始聚类中心，对176个像素点都进行同样的处理，形成若干个新聚类。For example: the value of N is 200, 24 equally divided points are respectively used as the initial clustering center, and the color Euclidean distance between 176 pixels and the initial clustering center is calculated, that is, 24 color Euclidean distances are calculated for each pixel , find the minimum Euclidean distance from the 24 color Euclidean distances, assign the A pixel to the initial cluster center corresponding to the minimum Euclidean distance, and perform the same processing on all 176 pixel points to form several new clusters .

3）计算每个新聚类的聚类中心，不断重复步骤2）直到每个聚类中心不再变化。3) Calculate the cluster center of each new cluster, and repeat step 2) until each cluster center does not change.

其中，聚类算法的目标是找到数据中自然聚类的中心，从而使得各个群组内部的均方误差总和J达到最小，即使得下式最小化：Among them, the goal of the clustering algorithm is to find the center of natural clustering in the data, so that the sum J of the mean square error within each group can be minimized, that is, the following formula can be minimized:

$\mathrm{<span><span>J</span>J</span>}<span><span>=</span>=</span>\underset{\mathrm{<span><span>n</span>no</span>}<span><span>=</span>=</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>N</span>N</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}\underset{\mathrm{<span><span>k</span>k</span>}<span><span>=</span>=</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>K</span>K</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}{\mathrm{<span><span>r</span>r</span>}}_{\mathrm{<span><span>nk</span>nk</span>}}<span><span>|</span>|</span><span><span>|</span>|</span>{\mathrm{<span><span>x</span>x</span>}}_{\mathrm{<span><span>n</span>no</span>}}<span><span>-</span>-</span>{\mathrm{<span><span>\&mu;</span>\&mu;</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>|</span>|</span>{<span><span>|</span>|</span>}^{\mathrm{<span><span>2</span>2</span>}}<span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>1</span>1</span>}<span><span>)</span>)</span>$

其中，r_nk用于说明第n个像素点是否被归类到第k类中，当像素点n被归类到第k类时r_nk=1；否则r_nk=0。x_n为第n个像素点的值，μ_k为第k类中像素点的平均值。本发明选择图像中像素颜色值的K等分点作为初始聚类中心，K为聚类数目。Among them, r_nk is used to indicate whether the n-th pixel is classified into the k-th class, when the pixel n is classified into the k-th class, r_nk =1; otherwise, r_nk =0. x_n is the value of the nth pixel, and μ_k is the average value of the kth pixel. The present invention selects K equal division points of pixel color values in the image as initial cluster centers, and K is the number of clusters.

103：对分割后阴影图像区域进行合并，得到最终分割结果；103: Merge the segmented shadow image regions to obtain a final segmentation result;

由于对预处理后阴影图像进行初始分割时仅利用了图像的颜色信息，即只是将颜色相近的像素点划分到同一类别，该方法通常得到的是一个过分割的结果，而不是对图像的精确分割。一幅图像由图像中的不同区域组成，区域的数量是有限的，过分割不仅可能导致阴影检测的不准确，而且会使检测时间增加。Since only the color information of the image is used for the initial segmentation of the preprocessed shadow image, that is, only the pixels with similar colors are classified into the same category, this method usually obtains an over-segmented result rather than an accurate image segmentation. segmentation. An image is composed of different areas in the image, and the number of areas is limited. Over-segmentation may not only lead to inaccurate shadow detection, but also increase the detection time.

1）比较分割后各相邻阴影图像区域的特征信息，当两个相互独立的区域满足颜色距离相近、边缘距离相近(即空间相邻)，并且区域连接处没有边缘时，对这两个区域进行合并；1) Compare the feature information of each adjacent shadow image area after segmentation. When two independent areas meet the similar color distance and edge distance (that is, the space is adjacent), and there is no edge at the area connection, the two areas to merge;

该方法使得被分割的阴影图像区域更加完整，同时避免了图像中阴影区域与亮度较低的非阴影区域的混淆，保证了分割结果的完整性和可靠性。This method makes the segmented shadow image area more complete, avoids the confusion of the shadow area in the image and the non-shadow area with low brightness, and ensures the integrity and reliability of the segmentation result.

初始分割后各阴影图像区域间的颜色距离定义为：The color distance between each shaded image region after the initial segmentation is defined as:

$\mathrm{<span><span>D</span>D.</span>}<span><span>\_</span>\_</span>{\mathrm{<span><span>color</span>color</span>}}_{\mathrm{<span><span>ij</span>ij</span>}}<span><span>=</span>=</span>\frac{<span><span>|</span>|</span>{\mathrm{<span><span>n</span>no</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>|</span>|</span><span><span>*</span>*</span><span><span>|</span>|</span>{\mathrm{<span><span>n</span>no</span>}}_{\mathrm{<span><span>j</span>j</span>}}<span><span>|</span>|</span>}{<span><span>|</span>|</span>{\mathrm{<span><span>n</span>no</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>|</span>|</span><span><span>+</span>+</span><span><span>|</span>|</span>{\mathrm{<span><span>n</span>no</span>}}_{\mathrm{<span><span>j</span>j</span>}}<span><span>|</span>|</span>}<span><span>|</span>|</span><span><span>|</span>|</span>{\stackrel{<span><span>\&OverBar;</span>\&OverBar;</span>}{\mathrm{<span><span>\&mu;</span>\&mu;</span>}}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>-</span>-</span>{\stackrel{<span><span>\&OverBar;</span>\&OverBar;</span>}{\mathrm{<span><span>\&mu;</span>\&mu;</span>}}}_{\mathrm{<span><span>j</span>j</span>}}<span><span>|</span>|</span><span><span>|</span>|</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>2</span>2</span>}<span><span>)</span>)</span>$

边缘距离为：The edge distance is:

D_edge_ij=||a_i-a_j||   (3)D_edge_ij =||a_i -a_j || (3)

其中，n_i表示第i个区域包含的像素个数，n_j表示第j个区域包含的像素个数，

与

分别表示第i和第j区域的颜色均值，a_i与a_j表示第i和第j区域边缘像素的均值。Among them, n_i represents the number of pixels contained in the i-th region, n_j represents the number of pixels contained in the j-th region,

and

Represent the color mean values of the i-th and j-th regions respectively, a_i and a_j represent the mean values of the edge pixels of the i-th and j-th regions.

根据实际应用中的需要设定颜色距离阈值和边缘距离阈值，选择出符合阈值的独立区域。Set the color distance threshold and the edge distance threshold according to the needs in practical applications, and select independent regions that meet the threshold.

对中值滤波后的阴影图像使用Canny算子检测边缘，当区域连接处与图像边缘有较多重合时，区域连接处两侧的区域不能进行合并。本发明中当区域连接处与边缘的重合像素数大于10时，不能对区域连接处两侧的区域进行合并。Use the Canny operator to detect the edge of the shadow image after the median filter. When the region connection overlaps with the image edge more, the regions on both sides of the region connection cannot be merged. In the present invention, when the number of overlapping pixels between the region junction and the edge is greater than 10, the regions on both sides of the region junction cannot be merged.

2）在某一区域合并前后，图像区域的颜色散度变化大于散度阈值时，自动停止区域合并过程，得到最终分割结果。2) Before and after a region is merged, when the color divergence change of the image region is greater than the divergence threshold, the region merge process is automatically stopped to obtain the final segmentation result.

本发明中，通过合并后的颜色散度与区域数目的比值T来判定是否停止区域合并。经过实验发现，当T≥0.05时，合并前后颜色散度有显著变化，需要停止区域合并过程，故设定阈值T₀=0.05。In the present invention, it is determined whether to stop region merging according to the ratio T of the combined color divergence to the number of regions. It is found through experiments that when T≥0.05, the color divergence changes significantly before and after merging, and it is necessary to stop the region merging process, so the threshold T₀ =0.05 is set.

104：对分割后的各个区域提取特征值，并获取特征向量；104: Extract feature values from each segmented region, and obtain feature vectors;

本发明对分割后的各个区域提取特征值，通过比较分割后各个区域组成的成对区域的4种特征值，即颜色直方图与纹理直方图、亮度比值、色差角度以及区域在图像中的距离，实现对原始图像的特征提取和阴影区域的自动分类。The present invention extracts eigenvalues for each segmented region, and compares the 4 kinds of eigenvalues of the paired regions formed by each segmented region, that is, the color histogram and texture histogram, brightness ratio, color difference angle, and the distance of the region in the image , to realize the feature extraction of the original image and the automatic classification of the shadow area.

基于成对区域的方法是指：对图像分割后的每个区域，与除去其自身外的所有其他区域进行配对，然后对这些配对后的成对区域进行比较的方法。The method based on paired regions refers to the method of pairing each segmented region of the image with all other regions except itself, and then comparing these paired paired regions.

颜色直方图与纹理直方图作为2个特征向量。其中，颜色直方图描述的是不同色彩在整幅图像中所占的比例，纹理直方图描述的是图像中灰度幅度的变化。Color histogram and texture histogram as 2 feature vectors. Among them, the color histogram describes the proportion of different colors in the whole image, and the texture histogram describes the change of gray scale in the image.

根据式(5)，判断R、G、B三个通道的平均亮度比值，包含3个特征向量：According to formula (5), the average brightness ratio of the three channels of R, G, and B is judged, including three feature vectors:

${\mathrm{<span><span>P</span>P</span>}}_{\mathrm{<span><span>R</span>R</span>}}<span><span>=</span>=</span>\frac{{\mathrm{<span><span>R</span>R</span>}}_{\mathrm{<span><span>avg</span>avg</span>}\mathrm{<span><span>1</span>1</span>}}}{{\mathrm{<span><span>R</span><figure-callout\; label="R\; avg"\; filenames="HDA00003181746200022.png"\; state="\{\{state\}\}">R</figure-callout></span><figure-callout\; label="R\; avg"\; filenames="HDA00003181746200022.png"\; state="\{\{state\}\}"></figure-callout>}}_{\mathrm{<span><span></span></span>}}}<span><span>,</span>,</span>$${\mathrm{<span><span>P</span>P</span>}}_{\mathrm{<span><span>G</span>G</span>}}<span><span>=</span>=</span>\frac{{\mathrm{<span><span>G</span>G</span>}}_{\mathrm{<span><span>avg</span>avg</span>}\mathrm{<span><span>1</span>1</span>}}}{{\mathrm{<span><span>G</span><figure-callout\; label="G\; avg"\; filenames="HDA00003181746200022.png"\; state="\{\{state\}\}">G</figure-callout></span><figure-callout\; label="G\; avg"\; filenames="HDA00003181746200022.png"\; state="\{\{state\}\}"></figure-callout>}}_{\mathrm{<span><span></span></span>}}}<span><span>,</span>,</span>$${\mathrm{<span><span>P</span>P</span>}}_{\mathrm{<span><span>B</span>B</span>}}<span><span>=</span>=</span>\frac{{\mathrm{<span><span>B</span>B</span>}}_{\mathrm{<span><span>avg</span>avg</span>}\mathrm{<span><span>1</span>1</span>}}}{{\mathrm{<span><span>B</span>B</span>}}_{\mathrm{<span><span>avg</span>avg</span>}\mathrm{<span><span>2</span>2</span>}}}<span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>4</span>4</span>}<span><span>)</span>)</span>$

其中，R_avg1代表成对区域中第1个区域的红色通道平均值，R_avg2代表成对区域中第2个区域的红色通道平均值，G_avg1代表成对区域中第1个区域的绿色通道平均值，G_avg2代表成对区域中第2个区域的绿色通道平均值，B_avg1代表成对区域中第1个区域的蓝色通道平均值，B_avg2代表成对区域中第2个区域的蓝色通道平均值。Among them, R_avg1 represents the average value of the red channel of the first region in the paired region, R_avg2 represents the average value of the red channel of the second region in the paired region, and G_avg1 represents the green channel of the first region in the paired region Average value, G_avg2 represents the average value of the green channel of the second region in the paired region, B_avg1 represents the average value of the blue channel of the first region in the paired region, B_avg2 represents the average value of the second region in the paired region Blue channel average.

同时，采用色差角度P_R/P_G和P_R/P_B作为2个特征向量，用于描述阴影区域与非阴影区域颜色分量比值的变化。At the same time, the color difference angles P_R /_PG and P_R /P_B are used as two feature vectors to describe the change of the ratio of the color components of the shaded area to the non-shaded area.

距离较远的区域对通常是由不同纹理构成的，因此，本发明将区域位置的归一化距离作为一个特征值，由区域面积的几何平均数除以区域中心的欧几里德距离得到1个特征向量，即：The region pair with a long distance is usually composed of different textures. Therefore, the present invention regards the normalized distance of the region position as a feature value, and divides the geometric mean of the region area by the Euclidean distance of the region center to obtain 1 eigenvectors, namely:

$\sqrt{{\mathrm{<span><span>s</span>the\; s</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>\&times;</span>\&times;</span>{\mathrm{<span><span>s</span>the\; s</span>}}_{\mathrm{<span><span>j</span>j</span>}}}<span><span>/</span>/</span>\mathrm{<span><span>d</span>d</span>}<span><span>-</span>-</span><span><span>-</span>-</span><span><span>-</span>-</span><span><span>(</span>(</span>\mathrm{<span><span>5</span>5</span>}<span><span>)</span>)</span>$

其中，s_i和s_j分别是第i和第j区域的面积，d是第i和第j区域中心的欧几里德距离。Among them, s_i and s_j are the areas of the i-th and j-th regions, respectively, and d is the Euclidean distance between the centers of the i-th and j-th regions.

具有相同反射比的成对区域，在相同光照条件下通常具有相似的纹理特征和色彩分布，但在不同光照条件下，则会具有相似的纹理特征和不同的颜色和亮度。因此，图像中具有相似颜色和纹理信息的区域，即颜色直方图、纹理直方图的卡方值较小的成对区域，往往具有相同光照强度。本发明通过计算R、G、B三个通道的平均亮度比值、阴影区域与非阴影区域颜色分量比值、区域位置的归一化距离，以及成对区域的颜色、纹理直方图的卡方值等4种特征值，即得到8个特征向量，将归一化后的特征向量作为SVM的输入量。Paired regions with the same reflectance usually have similar texture characteristics and color distribution under the same lighting conditions, but under different lighting conditions, they will have similar texture characteristics and different colors and brightness. Therefore, areas with similar color and texture information in the image, that is, paired areas with smaller chi-square values of the color histogram and texture histogram, tend to have the same light intensity. The present invention calculates the average brightness ratio of the three channels of R, G, and B, the color component ratio of the shaded area and the non-shaded area, the normalized distance of the area position, and the color of the paired area, the chi-square value of the texture histogram, etc. 4 kinds of eigenvalues, that is, 8 eigenvectors are obtained, and the normalized eigenvectors are used as the input of SVM.

105：对特征向量进行归一化处理，建立阴影区域分类模型；105: Perform normalization processing on the feature vector, and establish a shadow area classification model;

该步骤采用高斯径向基核函数作为支持向量机核函数，将归一化后的特征向量作为SVM的输入量，在阴影区域与非阴影区域两类之间寻找最优分类面，建立阴影区域分类模型。In this step, the Gaussian radial basis kernel function is used as the support vector machine kernel function, and the normalized feature vector is used as the input of SVM to find the optimal classification surface between the shaded area and the non-shaded area to establish the shaded area. classification model.

为了利用SVM实现阴影区域的自动检测，本发明首先选取一部分经过中值滤波处理的阴影图像作为训练集，对训练集中的所有样本，手动标注阴影区域、成对区域的关系，计算特征向量，进行归一化处理后，采用高斯径向基核函数作为支持向量机核函数，建立SVM的最优分类模型，最后利用SVM实现对未知图像中阴影和非阴影区域的自动检测。In order to use SVM to realize automatic detection of shadow areas, the present invention first selects a part of shadow images processed by median filtering as a training set, manually marks the relationship between shadow areas and paired areas for all samples in the training set, calculates feature vectors, and performs After normalization, the Gaussian radial basis kernel function is used as the support vector machine kernel function to establish the optimal classification model of SVM. Finally, the SVM is used to realize the automatic detection of shadow and non-shadow areas in unknown images.

高斯径向基核函数如式(6)所示：The Gaussian radial basis kernel function is shown in formula (6):

K(||x-x_c||)=exp(-||x-x_c||²/2σ²)   (6)K(||xx_c ||)=exp(-||xx_c ||² /2σ² ) (6)

其中，x为训练集的特征值，x_c为核函数中心，σ为函数的宽度参数，控制了函数的径向作用范围。Among them, x is the eigenvalue of the training set, x_c is the center of the kernel function, and σ is the width parameter of the function, which controls the radial range of the function.

106：对未知的自然场景阴影图像样本重复步骤101～104，提取未知图像特征值并作为支持向量机的输入量，根据阴影区域分类模型对图像区域进行判决，将不同区域分为阴影区域或非阴影区域。106: Repeat steps 101-104 for the shadow image sample of the unknown natural scene, extract the feature value of the unknown image and use it as the input of the support vector machine, judge the image area according to the shadow area classification model, and classify different areas into shadow areas or non-shaded areas. shaded area.

该步骤具体为：重复步骤101-104，对未知的自然场景阴影图像样本提取上述四个特征值，即提取表征颜色直方图与纹理直方图、亮度比值、色差角度以及区域在图像中的距离的特征值，作为未知图像特征值，并将未知图像特征值作为支持向量机的输入量，根据阴影区域分类模型，依据未知图像特征值对图像区域进行判决(即将每个未知图像特征值归类到阴影区域分类模型中)，将不同区域分为阴影区域或非阴影区域，实现阴影与非阴影区域的自动检测。This step is specifically: repeating steps 101-104, extracting the above four eigenvalues for unknown natural scene shadow image samples, that is, extracting the characteristic values of the color histogram and texture histogram, brightness ratio, color difference angle and the distance of the region in the image Eigenvalue, as the unknown image eigenvalue, and the unknown image eigenvalue as the input of the support vector machine, according to the shadow area classification model, judge the image area according to the unknown image eigenvalue (that is, classify each unknown image eigenvalue into In the shadow area classification model), different areas are divided into shadow areas or non-shade areas, and automatic detection of shadow and non-shade areas is realized.

下面以具体的实验来验证本方法提供的一种基于成对区域的单幅图像阴影检测方法的可行性，详见下文描述：The following specific experiments are used to verify the feasibility of a single image shadow detection method based on paired regions provided by this method. See the description below for details:

本发明选取64幅阴影图像，包括室内和室外场景，以及不同光照条件下的物体图像。其中，随机选取32幅阴影图像组成实验训练集，手动标注阴影区域以及成对区域的关系；另外32幅图像用于测试。The present invention selects 64 shadow images, including indoor and outdoor scenes, and object images under different lighting conditions. Among them, 32 shadow images are randomly selected to form the experimental training set, and the relationship between shadow areas and paired areas is manually marked; the other 32 images are used for testing.

参见图1，本发明以5幅自然场景阴影图像为例，受阴影区域中亮度变化、纹理信息、阴影目标与环境颜色相近等干扰因素影响，利用传统方法容易出现漏检的现象，从而导致阴影检测区域不完整（如图1(a)中人的手臂，图1(c)中木桩区域，图1(e)中人的颈部）。同时，对于亮度较低的非阴影区域，传统方法还存在一定的误检现象(如图1(a)中树丛区域，图1(c)中木桩的部分区域)。参见图2和图3，与传统方法相比，本发明克服了亮度变化、纹理信息等干扰因素的影响，有效减少了漏检和误检现象，阴影检测结果与真实情况基本一致，取得了比较理想的检测效果。Referring to Figure 1, the present invention takes 5 shadow images of natural scenes as an example. Due to the influence of interference factors such as brightness changes in the shadow area, texture information, shadow objects and environmental colors, it is easy to miss detection by using traditional methods, resulting in shadows. The detection area is incomplete (such as the human arm in Figure 1(a), the stake area in Figure 1(c), and the human neck in Figure 1(e)). At the same time, for the non-shaded areas with low brightness, the traditional method still has some false detections (such as the bush area in Figure 1(a), and some areas of wooden piles in Figure 1(c)). Referring to Fig. 2 and Fig. 3, compared with the traditional method, the present invention overcomes the influence of interference factors such as brightness change and texture information, effectively reduces the phenomenon of missed detection and false detection, and the shadow detection result is basically consistent with the real situation, achieving a comparative advantage. Ideal detection effect.

本发明以图1的5幅原始自然场景阴影图像为例，分别利用本发明的聚类与区域合并算法，以及传统方法采用的均值漂移算法，实际分割区域数目如表1所示。The present invention takes the five original natural scene shadow images in Fig. 1 as an example, uses the clustering and region merging algorithm of the present invention, and the mean shift algorithm adopted by the traditional method respectively, and the actual number of segmented regions is shown in Table 1.

表1Table 1

利用传统方法与本发明对图1中5幅图像进行阴影检测，运行时间如表2所示：Utilize traditional method and the present invention to carry out shadow detection to 5 images in Fig. 1, running time is as shown in table 2:

表2Table 2

从表1、2可以看出，传统方法中的均值漂移算法对局部细节、噪声比较敏感，对于具有复杂纹理信息的自然场景图像容易产生严重的过分割，这会大大增加支持向量机的测试时间；而本发明实现了大多数区域的合并，有效减少了分割区域的数目，大大提高了算法的运行速度。It can be seen from Tables 1 and 2 that the mean shift algorithm in the traditional method is sensitive to local details and noise, and is prone to severe over-segmentation for natural scene images with complex texture information, which will greatly increase the test time of the support vector machine ; while the present invention realizes the merging of most areas, effectively reduces the number of divided areas, and greatly improves the running speed of the algorithm.

参见图4，提取特征向量，并结合SVM进行分类测试。根据实验训练集建立具有相同反射比的成对区域的模型，以及具有不同反射比的成对区域的模型，使用SVM对模型寻找最优参数，实现图像中各区域的最高分类正确率。寻优过程中，两个参数不断变化形成分类正确率的等值曲线，图4中曲线上的数值显示了各个参数对应的分类正确率。图4中纵坐标表示SVM高斯径向基核函数中参数g的对数值，横坐标表示惩罚参数c的对数值。经过SVM参数寻优，最终实现94.17%的分类正确率，此时两个参数的取值分别为：c=16，g=2，即log₂c=4，log₂g=1。Referring to Figure 4, feature vectors are extracted and combined with SVM for classification testing. Based on the experimental training set, a model of paired regions with the same reflectance and a model of paired regions with different reflectances are established, and SVM is used to find the optimal parameters for the model to achieve the highest classification accuracy of each region in the image. During the optimization process, the two parameters are constantly changing to form the equivalent curve of the classification accuracy rate. The values on the curve in Figure 4 show the classification accuracy rate corresponding to each parameter. In Figure 4, the ordinate represents the logarithmic value of the parameter g in the SVM Gaussian radial basis kernel function, and the abscissa represents the logarithmic value of the penalty parameter c. After optimizing the SVM parameters, a classification accuracy rate of 94.17% was finally achieved. At this time, the values of the two parameters are: c=16, g=2, that is, log₂ c=4, log₂ g=1.

本发明实施例方法中应用到的中值滤波算法、聚类算法和SVM算法等均为数据处理方法中的公知技术，本发明实施例在此不做赘述。The median filter algorithm, clustering algorithm, and SVM algorithm used in the method of the embodiment of the present invention are all known technologies in the data processing method, and are not described in detail in the embodiment of the present invention.

本领域技术人员可以理解附图只是一个优选实施例的示意图，上述本发明实施例序号仅仅为了描述，不代表实施例的优劣。Those skilled in the art can understand that the accompanying drawing is only a schematic diagram of a preferred embodiment, and the serial numbers of the above-mentioned embodiments of the present invention are for description only, and do not represent the advantages and disadvantages of the embodiments.

以上所述仅为本发明的较佳实施例，并不用以限制本发明，凡在本发明的精神和原则之内，所作的任何修改、等同替换、改进等，均应包含在本发明的保护范围之内。The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection of the present invention. within range.

Claims (5)

Translated fromChinese

1.一种基于成对区域的单幅图像阴影检测方法，其特征在于，所述方法包括以下步骤：1. a single image shadow detection method based on paired regions, is characterized in that, described method comprises the following steps:(1)对已知的自然场景阴影图像利用中值滤波方法进行预处理，获取预处理后阴影图像；(1) Preprocessing the known natural scene shadow image by using the median filter method to obtain the preprocessed shadow image;(2)利用聚类方法对所述预处理后阴影图像进行初始分割，获取分割后阴影图像区域；(2) Utilize the clustering method to carry out initial segmentation to the shadow image after the pretreatment, obtain the shadow image area after segmentation;(3)对所述分割后阴影图像区域进行合并，得到最终分割结果；(3) merging the shadow image regions after the segmentation to obtain the final segmentation result;(4)对分割后的各个区域提取特征值，并获取特征向量；(4) Extract feature values for each segmented region, and obtain feature vectors;(5)对特征向量进行归一化处理，建立阴影区域分类模型；(5) Normalize the eigenvectors to establish a shadow area classification model;(6)对未知的自然场景阴影图像样本重复步骤(1)～(4)，提取未知图像特征值并作为支持向量机的输入量，根据所述阴影区域分类模型对图像区域进行判决，将不同区域分为阴影区域或非阴影区域。(6) Repeat steps (1) to (4) for the unknown natural scene shadow image sample, extract the unknown image feature value and use it as the input of the support vector machine, judge the image area according to the shadow area classification model, and classify different Areas are classified as shaded or unshaded.2.根据权利要求1所述的一种基于成对区域的单幅图像阴影检测方法，其特征在于，所述利用聚类方法对所述预处理后阴影图像进行初始分割，获取分割后阴影图像区域的步骤具体包括：2. a kind of single image shadow detection method based on paired regions according to claim 1, is characterized in that, described utilizes clustering method to carry out initial segmentation to described preprocessed shadow image, obtains the shadow image after segmentation The regional steps specifically include:1)将所述预处理后阴影图像中所有像素划分为K个组；1) dividing all pixels in the shadow image after the preprocessing into K groups;2)选择图像中所有像素颜色值的K个等分点作为初始聚类中心，计算N-K个像素点与初始聚类中心的颜色欧氏距离，将N-K个像素点分别分配给欧式距离最小的聚类，形成新聚类；2) Select K equally divided points of all pixel color values in the image as the initial cluster center, calculate the color Euclidean distance between the N-K pixel points and the initial cluster center, and assign the N-K pixel points to the cluster with the smallest Euclidean distance. class to form a new cluster;3)计算每个所述新聚类的聚类中心，不断重复步骤2)直到每个聚类中心不再变化。3) Calculate the cluster centers of each of the new clusters, and repeat step 2) until each cluster center does not change.3.根据权利要求1所述的一种基于成对区域的单幅图像阴影检测方法，其特征在于，所述对所述分割后阴影图像区域进行合并，得到最终分割结果的步骤具体包括：3. A method for detecting shadows of a single image based on paired regions according to claim 1, wherein the step of merging the shadow image regions after the segmentation to obtain the final segmentation result specifically includes:1)比较分割后各相邻阴影图像区域的特征信息，当两个相互独立的区域满足颜色距离阈值、边缘距离阈值，并且区域连接处没有边缘时，对这两个区域进行合并；1) Compare the feature information of each adjacent shadow image region after segmentation, and when two mutually independent regions meet the color distance threshold and edge distance threshold, and there is no edge at the region connection, merge the two regions;2)在某一区域合并前后，图像区域的颜色散度变化大于散度阈值时，自动停止区域合并过程，得到所述最终分割结果。2) Before and after a region is merged, when the change of the color divergence of the image region is greater than the divergence threshold, the region merge process is automatically stopped, and the final segmentation result is obtained.4.根据权利要求1所述的一种基于成对区域的单幅图像阴影检测方法，其特征在于，所述对分割后的各个区域提取特征值，并获取特征向量的步骤具体包括：4. A method for detecting shadows in a single image based on paired regions according to claim 1, wherein the step of extracting eigenvalues from each segmented region and obtaining a eigenvector specifically comprises:比较分割后各个区域组成的成对区域的4种特征值，即颜色直方图与纹理直方图、亮度比值、色差角度以及区域在图像中的距离，通过4种特征值得到8个特征向量；其中，颜色直方图与纹理直方图作为2个特征向量；R、G、B三个通道的平均亮度比值，作为3个特征向量：Compare the 4 kinds of eigenvalues of the paired regions composed of each region after segmentation, namely the color histogram and texture histogram, the brightness ratio, the color difference angle and the distance of the region in the image, and obtain 8 eigenvectors through the 4 kinds of eigenvalues; , the color histogram and texture histogram are used as two feature vectors; the average brightness ratio of the three channels of R, G, and B is used as three feature vectors:${\mathrm{<span><span>P</span>P</span>}}_{\mathrm{<span><span>R</span>R</span>}}<span><span>=</span>=</span>\frac{{\mathrm{<span><span>R</span>R</span>}}_{\mathrm{<span><span>avg</span>avg</span>}\mathrm{<span><span>1</span>1</span>}}}{{\mathrm{<span><span>R</span>R</span>}}_{\mathrm{<span><span>avg</span>avg</span>}\mathrm{<span><span>2</span>2</span>}}}<span><span>,</span>,</span>$${\mathrm{<span><span>P</span>P</span>}}_{\mathrm{<span><span>G</span>G</span>}}<span><span>=</span>=</span>\frac{{\mathrm{<span><span>G</span>G</span>}}_{\mathrm{<span><span>avg</span>avg</span>}\mathrm{<span><span>1</span>1</span>}}}{{\mathrm{<span><span>G</span>G</span>}}_{\mathrm{<span><span>avg</span>avg</span>}\mathrm{<span><span>2</span>2</span>}}}<span><span>,</span>,</span>$${\mathrm{<span><span>P</span>P</span>}}_{\mathrm{<span><span>B</span>B</span>}}<span><span>=</span>=</span>\frac{{\mathrm{<span><span>B</span>B</span>}}_{\mathrm{<span><span>avg</span>avg</span>}\mathrm{<span><span>1</span>1</span>}}}{{\mathrm{<span><span>B</span>B</span>}}_{\mathrm{<span><span>avg</span>avg</span>}\mathrm{<span><span>2</span>2</span>}}}$其中，R_avg1代表成对区域中第1个区域的红色通道平均值，R_avg2代表成对区域中第2个区域的红色通道平均值，G_avg1代表成对区域中第1个区域的绿色通道平均值，G_avg2代表成对区域中第2个区域的绿色通道平均值，B_avg1代表成对区域中第1个区域的蓝色通道平均值，B_avg2代表成对区域中第2个区域的蓝色通道平均值；Among them, R_avg1 represents the average value of the red channel of the first region in the paired region, R_avg2 represents the average value of the red channel of the second region in the paired region, and G_avg1 represents the green channel of the first region in the paired region Average value, G_avg2 represents the average value of the green channel of the second region in the paired region, B_avg1 represents the average value of the blue channel of the first region in the paired region, B_avg2 represents the average value of the second region in the paired region blue channel average;采用色差角度P_R/P_G和P_R/P_B作为2个特征向量，由区域面积的几何平均数除以区域中心的欧几里德距离得到1个特征向量，即：Using the color difference angles P_R /_PG and P_R /P_B as two eigenvectors, one eigenvector is obtained by dividing the geometric mean of the area area by the Euclidean distance of the area center, namely:$\sqrt{{\mathrm{<span><span>s</span>the\; s</span>}}_{\mathrm{<span><span>i</span>i</span>}}<span><span>\&times;</span>\&times;</span>{\mathrm{<span><span>s</span>the\; s</span>}}_{\mathrm{<span><span>j</span>j</span>}}}<span><span>/</span>/</span>\mathrm{<span><span>d</span>d</span>}$其中，s_i和s_j分别是第i和第j区域的面积，d是第i和第j区域中心的欧几里德距离。Among them, s_i and s_j are the areas of the i-th and j-th regions, respectively, and d is the Euclidean distance between the centers of the i-th and j-th regions.5.根据权利要求1所述的一种基于成对区域的单幅图像阴影检测方法，其特征在于，所述对特征向量进行归一化处理，建立阴影区域分类模型的步骤具体包括：5. A method for detecting shadows of a single image based on paired regions according to claim 1, wherein the step of normalizing the feature vectors and establishing a shadow region classification model specifically includes:选取一部分经过中值滤波处理的阴影图像作为训练集，对训练集中的所有样本，手动标注阴影区域、成对区域的关系，计算特征向量，将归一化后的特征向量作为SVM的输入量，采用高斯径向基核函数作为支持向量机核函数，建立SVM的最优分类模型。Select a part of the shadow image processed by the median filter as the training set, manually mark the relationship between the shadow area and the paired area for all samples in the training set, calculate the feature vector, and use the normalized feature vector as the input of the SVM, Gaussian radial basis kernel function is used as the kernel function of support vector machine to establish the optimal classification model of SVM.

Images