CN101251373A - Rapid detection method of microstructure three-dimensional stereo image - Google Patents

Abstract

Translated fromChinese

本发明涉及一种微结构三维尺寸立体图像快速检测方法，特别是涉及采用立体图像对建立匹配关系，通过物像反求模型计算微结构长度、宽度和高度的方法。该方法包括以下主要步骤：预处理图像、提取图像中的特征点、特征匹配、匹配结果优化和纠错、物像反求和输出三维尺寸。本发明属于非接触式测量，不破坏样品表面，而且能够保证系统的测量精度，同时，通过生成立体表面提高检测速度，操作流程更简捷，是一种效率很高的计算机辅助测量法。

The invention relates to a rapid detection method of a three-dimensional stereoscopic image of a microstructure, in particular to a method for establishing a matching relationship by using a stereoscopic image pair, and calculating the length, width and height of a microstructure through an object image reverse model. The method includes the following main steps: preprocessing the image, extracting feature points in the image, feature matching, optimization and error correction of matching results, reverse summation of object images and outputting three-dimensional dimensions. The invention belongs to non-contact measurement, does not damage the surface of the sample, and can ensure the measurement accuracy of the system. At the same time, the detection speed is improved by generating a three-dimensional surface, and the operation process is simpler. It is a computer-aided measurement method with high efficiency.

Description

Translated fromChinese

微结构三维尺寸立体图像快速检测方法Rapid detection method of microstructure three-dimensional stereo image

技术领域technical field

本发明涉及一种微结构三维尺寸立体图像检测方法，特别是涉及采用立体图像对建立匹配关系，通过物像反求模型计算微结构长度、宽度和高度的方法。The invention relates to a method for detecting a three-dimensional stereoscopic image of a microstructure, in particular to a method for establishing a matching relationship by using a stereoscopic image pair, and calculating the length, width and height of a microstructure through an object image inverse model.

背景技术Background technique

微结构三维尺寸检测是微机电系统中的重要方向。微结构的尺度范围在几微米-几十微米左右，体积很小，其三维尺寸估计是微结构加工和生产中必须的。微结构是一个比较大的范围，可以是微器件，如沟道、凸台等，还可以是由多个微器件中构成的组合系统。微结构一般要经过多道工序才能完成，受加工条件等外部因素的影响，在摸索加工工艺阶段，经常需要估计微结构的长度、宽度和高度，然后根据检测结果，不断调整工艺参数直到达到合理技术指标为止。Three-dimensional size detection of microstructure is an important direction in MEMS. The scale of the microstructure ranges from a few microns to tens of microns, and its volume is very small. Estimating its three-dimensional size is necessary in the processing and production of microstructures. Microstructure is a relatively large range, which can be micro-devices, such as channels, bosses, etc., or a combined system composed of multiple micro-devices. The microstructure generally needs to go through multiple processes to complete. Affected by external factors such as processing conditions, it is often necessary to estimate the length, width and height of the microstructure during the stage of exploring the processing technology, and then continuously adjust the process parameters according to the test results until a reasonable technical indicators.

目前，已出现了多种尺寸检测方法，扫描电子显微镜检测法和原子力显微镜法是常用的两种微结构尺寸检测方法，这两类方法检测精度较高，但属于接触式测量方法，对微结构表面有破坏作用。而且这两类方法的操作流程很复杂，完成一个测量过程需要花费较长的时间，不便于进行摸索加工工艺阶段的快速测量。At present, a variety of size detection methods have appeared. Scanning electron microscopy and atomic force microscopy are two commonly used microstructure size detection methods. These two types of methods have high detection accuracy, but they are contact measurement methods. The surface is destructive. Moreover, the operation procedures of these two types of methods are very complicated, and it takes a long time to complete a measurement process, which is not convenient for rapid measurement in the stage of exploring the processing technology.

干涉法、三角测量法、条纹投影法也用于了微结构三维尺寸的测量，这些方法属于非接触式测量方法，对微结构表面没有破坏作用，但是其检测分辨率和精度偏低，适合于尺度较大的微结构三维尺寸测量。这些方法对图像中的条纹间距、频率信息进行分析，检测精度受外部辅助光学系统的影响明显，操作流程较为复杂。Interferometry, triangulation, and fringe projection methods are also used to measure the three-dimensional dimensions of microstructures. These methods are non-contact measurement methods that do not damage the surface of microstructures, but their detection resolution and accuracy are low, and they are suitable for Three-dimensional size measurement of large-scale microstructures. These methods analyze the fringe spacing and frequency information in the image, the detection accuracy is obviously affected by the external auxiliary optical system, and the operation process is relatively complicated.

发明内容Contents of the invention

针对上述现有的微结构三维尺寸检测方法所存在的问题，本发明推出适合于小尺度的非接触式的微结构三维尺寸立体图像快速检测方法，其目的在于采用立体图像对建立匹配关系，通过物像反求模型计算微结构长度、宽度和高度，自动给出微结构兴趣区的长度、宽度和高度数值。In view of the problems existing in the above-mentioned existing microstructure three-dimensional size detection method, the present invention introduces a non-contact microstructure three-dimensional size stereoscopic image rapid detection method suitable for small scales. The purpose is to use stereoscopic image pairs to establish a matching relationship. The object image inverse model calculates the length, width and height of the microstructure, and automatically gives the length, width and height values of the microstructure ROI.

为了实现上述目的，本发明采取了如下技术方案。由光学体视显微镜和两个CCD摄像头组成的显微立体视觉系统采集立体图像对，并传送给计算机，由计算机对立体图像对进行处理和匹配，通过物像反求获得微结构的三维立体表面，从而确定兴趣区的长度、宽度和高度。所述的微结构三维尺寸立体图像检测方法包括以下步骤：In order to achieve the above object, the present invention adopts the following technical solutions. The microscopic stereo vision system consisting of an optical stereo microscope and two CCD cameras collects stereo image pairs and transmits them to the computer. The computer processes and matches the stereo image pairs, and obtains the three-dimensional surface of the microstructure through object image inversion. , so as to determine the length, width and height of the ROI. The three-dimensional image detection method of the microstructure comprises the following steps:

1、预处理图像1. Preprocess the image

向计算机输入左图像和右图像，然后由计算机进行滤波、目标分割处理，区分开图像中的目标像素和背景像素，目标像素设置为黑色，背景像素设置为白色，得到左右预处理图像和左右目标图像。Input the left image and right image to the computer, and then the computer performs filtering and target segmentation processing to distinguish the target pixels and background pixels in the image. The target pixels are set to black, and the background pixels are set to white to obtain the left and right preprocessed images and the left and right targets. image.

2、提取图像中的特征点像素2. Extract the feature point pixels in the image

针对左目标图像和右目标图像，采用通用的特征提取方法提取左图像和右图像中的特征点像素，记录左图像中特征点像素的图像坐标和右图像中特征点像素的图像坐标。For the left target image and the right target image, a common feature extraction method is used to extract the feature point pixels in the left image and the right image, and the image coordinates of the feature point pixels in the left image and the image coordinates of the feature point pixels in the right image are recorded.

采用的特征提取方法包括求局部最大熵法、小波变换法、Sober算子法和灰度阈值法。The feature extraction methods used include local maximum entropy method, wavelet transform method, Sober operator method and gray threshold method.

3、特征匹配3. Feature matching

以左图像中的特征点像素为种子像素，在右图像中设置对应的搜索区域，搜索区域设置方法为：若种子像素在左图像中的图像坐标为(w₁^l，w₂^l)，在右图像中以坐标(w₁^l，w₂^l)对应的像素为中心，长方形区域[w₁^l±d₁，w₂^l±d₂]即为搜索区域，其中1≤d₁≤10，0≤d₂≤4。The feature point pixels in the left image are used as the seed pixels, and the corresponding search area is set in the right image. The search area setting method is: if the image coordinates of the seed pixel in the left image are (w₁^l , w₂^l ), in In the right image, the pixel corresponding to the coordinates (w₁^l , w₂^l ) is the center, and the rectangular area [w₁^l ±d₁ , w₂^l ±d₂ ] is the search area, where 1≤d₁ ≤10,_0≤d2≤4 .

遍历搜索区域中的所有相关点像素，计算种子像素和相关点像素之间的相似度，得到搜索区域对应的相似度集合，对相似度集合进行统计，获得搜索区域对应的区域相似度。统计方法包括求最大相似度的方法，对相似度集合求平均的方法。Traverse all relevant point pixels in the search area, calculate the similarity between the seed pixel and the relevant point pixel, obtain the similarity set corresponding to the search area, and perform statistics on the similarity set to obtain the regional similarity corresponding to the search area. Statistical methods include the method of seeking the maximum similarity and the method of averaging the similarity sets.

把区域相似度与设定的相似度阈值比较，判断该搜索区域内是否存在匹配点像素，如果区域相似度大于相似度阈值，则当前种子像素的匹配点像素存在，否则匹配点像素不存在。遍历左图像中的所有种子像素，依据相同的搜索策略，在右图像中找出所有种子像素对应的匹配点像素。相似度阈值取区间[0.6，0.9]内的值。Compare the regional similarity with the set similarity threshold to determine whether there is a matching point pixel in the search area. If the regional similarity is greater than the similarity threshold, the matching point pixel of the current seed pixel exists, otherwise the matching point pixel does not exist. Traversing all the seed pixels in the left image, according to the same search strategy, find the matching point pixels corresponding to all the seed pixels in the right image. The similarity threshold takes a value within the interval [0.6, 0.9].

4、匹配结果优化和纠错4. Matching result optimization and error correction

采用外极线约束条件和一致性约束原则对匹配结果进行优化和纠错，进一步排除匹配点像素集合中的错匹配点像素和误匹配点像素，输出匹配点像素的图像坐标。The matching result is optimized and corrected by using the epipolar line constraint condition and the consistency constraint principle, further eliminating the wrong matching point pixel and the wrong matching point pixel in the matching point pixel set, and outputting the image coordinates of the matching point pixel.

5、物像反求5. Reverse object image

以左右图像中的种子像素和对应匹配点像素的图像坐标及视差为自变量，采用计算机双目立体视觉原理中的物像反求模型，计算出所有种子像素和匹配点像素对应的物空间相关点的空间坐标集合。若某种子像素的图像坐标为(w₁^l，w₂^l)，对应的匹配点像素坐标为(w₁^r，w₂^r)，视差定义为w₁^l-w₁^r。采用的物像反求模型包括小孔成像模型、投影模型和弱视差显微立体视觉模型。Taking the image coordinates and parallax of the seed pixels in the left and right images and the corresponding matching point pixels as independent variables, using the object image inversion model in the principle of computer binocular stereo vision, the object space correlation corresponding to all seed pixels and matching point pixels is calculated. A collection of spatial coordinates of points. If the image coordinates of a certain sub-pixel are (w₁^l , w₂^l ), the pixel coordinates of the corresponding matching point are (w₁^r , w₂^r ), and the parallax is defined as w₁^l -w₁^r . The object image inverse model adopted includes pinhole imaging model, projection model and weak parallax microscopic stereoscopic vision model.

对物像反求模型获得的物空间相关点的坐标集合进行处理，采用图形学中的三角剖分法对三维空间点造型，得到空间曲面。The coordinate set of the object space related points obtained by the object image inverse model is processed, and the three-dimensional space point is modeled by the triangulation method in graphics to obtain the space surface.

6、输出三维尺寸6. Output 3D size

选择三维重构面的被测区域，输出该区域的长度、宽度和高度方向的距离，该距离即为被测的三维尺寸。Select the measured area of the three-dimensional reconstruction surface, and output the distance in the length, width and height directions of the area, and the distance is the measured three-dimensional size.

本发明所涉及的微结构三维尺寸快速检测方法采用了处理立体图像对建立匹配关系计算立体尺寸的方法，这种测量方式属于非接触式测量，不破坏样品表面，而且能够保证系统的测量精度，同时，通过生成立体表面提高检测速度，操作流程更简捷，是一种效率很高的计算机辅助测量法。The rapid three-dimensional size detection method of the microstructure involved in the present invention adopts the method of processing the three-dimensional image pair to establish a matching relationship and calculate the three-dimensional size. This measurement method belongs to non-contact measurement, does not damage the sample surface, and can ensure the measurement accuracy of the system. At the same time, the detection speed is improved by generating a three-dimensional surface, the operation process is simpler, and it is a highly efficient computer-aided measurement method.

附图说明Description of drawings

图1为本发明涉及的微结构三维尺寸立体图像快速检测方法示意图Fig. 1 is a schematic diagram of a method for rapid detection of microstructure three-dimensional stereoscopic images involved in the present invention

图2为本发明涉及的特征匹配方法示意图Fig. 2 is a schematic diagram of the feature matching method involved in the present invention

附图中标记说明Explanation of marks in the drawings

S11、输入左图像S11, input left image

S12、输入右图像S12. Input the right image

S21、预处理图像S21, preprocessing image

S31、识别左右图像中的目标S31. Recognize the target in the left and right images

S32、输出的左目标图像S32, the output left target image

S33、输出的右目标图像S33, the output right target image

S41、提取左图像特征S41, extracting left image features

S42、提取右图像特征S42, extracting right image features

S51、特征匹配S51. Feature matching

S52、匹配结果优化和纠错S52. Matching result optimization and error correction

S53、匹配点图像坐标输出S53, matching point image coordinate output

S54、左图像特征点S54, left image feature points

S55、计算种子和相关点的相似度S55, calculate the similarity between the seed and the related point

S56、遍历图像中的所有特征点S56, traversing all feature points in the image

S57、匹配结束S57, end of matching

S58、选择相关点S58. Select relevant points

S59、设置右图像匹配区域S59, setting the right image matching area

S591、存储相似度计算结果S591. Store the similarity calculation result

S592、判断是否遍历完匹配区域内的所有像素S592, judging whether all the pixels in the matching area have been traversed

S593、判断集合中最大相似度是否大于设置的阈值S593, judging whether the maximum similarity in the set is greater than a set threshold

S594、舍弃该种子像素S594. Discard the sub-pixel

S595、设置相关点为匹配点S595. Set the relevant points as matching points

S61、物像反求S61. Reverse object image

S71、生成的三维立体图S71, the generated three-dimensional stereogram

S81、测量区域的选择S81. Selection of measurement area

S91、输出三维尺寸：长度、宽度和高度S91. Outputting three-dimensional dimensions: length, width and height

具体实施方式Detailed ways

现结合附图对本发明作进一步详细阐述。图1和图2显示本发明涉及的微结构三维尺寸立体图像快速检测方法的流程图，如图所示，微结构三维尺寸立体图像快速检测方法包括以下步骤：The present invention is described in further detail now in conjunction with accompanying drawing. Fig. 1 and Fig. 2 show the flowchart of the rapid detection method of three-dimensional stereoscopic image of microstructure involved in the present invention, as shown in the figure, the rapid detection method of three-dimensional stereoscopic image of microstructure comprises the following steps:

1、预处理图像1. Preprocess the image

向计算机输入左图像S11和右图像S12，采用彩色图像向量中值滤波方法进行滤波，得到滤波后的彩色图像IAL和IAR。然后对彩色图像采用下面的方法分割目标和图像背景。The left image S11 and the right image S12 are input to the computer, and the color image vector median filter method is used for filtering to obtain filtered color images IAL and IAR. Then use the following method to segment the target and the image background for the color image.

(1)针对整幅彩色图像IAL(IAR)中的所有像素，将图像IAL(IAR)转化为灰度图像IBL(IBR)。转化方法为：灰度图像IBL(IBR)中的像素的红、绿、蓝灰度值相等，且等于图像IAL(IAR)中对应像素的红、绿、蓝灰度值的求和平均值。(1) For all pixels in the entire color image IAL(IAR), convert the image IAL(IAR) into a grayscale image IBL(IBR). The conversion method is: the red, green and blue gray values of the pixels in the grayscale image IBL (IBR) are equal, and equal to the summed average of the red, green and blue gray values of the corresponding pixels in the image IAL (IAR).

(2)针对灰度图像IBL(IBR)中的所有像素，将图像IBL(IBR)转化为分割图像ICL(ICR)，分割方法为：判断灰度图像IBL(IBR)中的像素灰度值是否大于设定的阈值TM，如果大于关系成立，则为图像ICL(ICR)中的背景像素，设置为白色，如果大于关系不成立，则为图像ICL(ICR)中的目标像素，设置为黑色，TM取100。(2) For all the pixels in the grayscale image IBL(IBR), convert the image IBL(IBR) into a segmented image ICL(ICR). The segmentation method is: determine whether the pixel gray value in the grayscale image IBL(IBR) is Greater than the set threshold TM, if the greater than the relationship is established, it is the background pixel in the image ICL (ICR), set to white, if the greater than the relationship is not established, then it is the target pixel in the image ICL (ICR), set to black, TM Take 100.

(3)对图像ICL(ICR)进行面积滤波，滤除伪目标像素，得到滤波后的图像IDL(IDR)。采用逐点搜索的方法，寻找图像ICL(ICR)中各孤立分布的黑色像素团，并记录各黑色像素团的黑色像素数目，如果该黑色像素团的黑色像素数目大于设定的阈值，该阈值取200，则为目标标记像素，否则为伪目标标记像素，设置为白色。(3) Perform area filtering on the image ICL (ICR) to filter out false target pixels to obtain the filtered image IDL (IDR). Use the method of point-by-point search to find each isolated black pixel group in the image ICL (ICR), and record the number of black pixels in each black pixel group. If the number of black pixels in the black pixel group is greater than the set threshold, the threshold If it is 200, it is a target marker pixel, otherwise it is a false target marker pixel, and it is set to white.

(4)图像颜色恢复。把图像IDL(IDR)中的黑色目标标记像素的颜色值设置为图像IAL(IAR)中的对应位置处的像素颜色值，图像IAL(IAR)即为预处理后的图像S32(S33)。(4) Image color restoration. The color value of the black target mark pixel in the image IDL (IDR) is set to the pixel color value at the corresponding position in the image IAL (IAR), and the image IAL (IAR) is the preprocessed image S32 (S33).

2、提取图像中的特征点像素2. Extract the feature point pixels in the image

针对输出的左目标图像S32和输出的右目标图像S33，采用求局部最大熵的特征提取方法，设目标图像S32和S33中的当前像素的灰度值为f(i，j)，(i，j)为当前像素的图像坐标，选择该像素的m×n邻域，m、n设置为8，计算该区域内的像素的熵值：For the left target image S32 of output and the right target image S33 of output, adopt the feature extraction method of seeking local maximum entropy, set the grayscale value of the current pixel in target image S32 and S33 f (i, j), (i, j) is the image coordinate of the current pixel, select the m×n neighborhood of the pixel, m and n are set to 8, and calculate the entropy value of the pixel in this area:

${\mathrm{<span><span>H</span>h</span>}}_{\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}}<span><span>=</span>=</span>\mathrm{<span><span>f</span>f</span>}<span><span>(</span>(</span>\mathrm{<span><span>i</span>i</span>}<span><span>,</span>,</span>\mathrm{<span><span>j</span>j</span>}<span><span>)</span>)</span><span><span>/</span>/</span>\underset{\mathrm{<span><span>k</span>k</span>}<span><span>=</span>=</span>\mathrm{<span><span>i</span>i</span>}<span><span>-</span>-</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>i</span>i</span>}<span><span>+</span>+</span>\mathrm{<span><span>1</span>1</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}\underset{\mathrm{<span><span>p</span>p</span>}<span><span>=</span>=</span>\mathrm{<span><span>j</span>j</span>}<span><span>-</span>-</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>j</span>j</span>}<span><span>+</span>+</span>\mathrm{<span><span>1</span>1</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}\mathrm{<span><span>f</span>f</span>}<span><span>(</span>(</span>\mathrm{<span><span>k</span>k</span>}<span><span>,</span>,</span>\mathrm{<span><span>p</span>p</span>}<span><span>)</span>)</span>$

H_i，j为(i，j)对应得8邻域的熵值，k、p为8领域内的像素的图像坐标，若H_i，j＞0.05，则为特征点像素，保持原来颜色，否则为背景像素，设置为白色，由此，得到提取的左图像特征点像素S41和提取的右图像特征点像素S42。H_{i, j} is the entropy value of (i, j) corresponding to 8 neighborhoods, k, p are the image coordinates of pixels in the 8 areas, if H_{i, j} > 0.05, it is a feature point pixel, keep the original color, Otherwise, it is a background pixel, which is set to white, thereby obtaining the extracted left image feature point pixel S41 and the extracted right image feature point pixel S42.

3、特征匹配3. Feature matching

针对提取的左图像特征像素S41和提取的右图像特征像素S42，以S41中的当前特征点为种子像素，在S42中对应位置处设置右图像匹配区域S59，设置方法为：若种子像素在左图像中的图像坐标为(w₁^l，w₂^l)，在右图像中以坐标(w₁^l，w₂^l)对应的像素为中心，长方形区域[w₁^l±10，w₂^l±10]作为搜索区域。For the extracted left image feature pixel S41 and the extracted right image feature pixel S42, with the current feature point in S41 as the seed pixel, the right image matching area S59 is set at the corresponding position in S42, and the setting method is: if the seed pixel is in the left The image coordinates in the image are (w₁^l , w₂^l ), in the right image, the pixel corresponding to the coordinates (w₁^l , w₂^l ) is the center, and the rectangular area [w₁^l ±10, w₂^l ± 10] as the search area.

选择S59中的相关点像素S58，遍历搜索区域中的所有相关点像素，计算种子像素和所有相关点像素之间的相似度，存储相似度计算结果S591，并构成右图像匹配区域S59的相似度集合，相似度计算使用下面的通用公式：Select the relevant point pixel S58 in S59, traverse all relevant point pixels in the search area, calculate the similarity between the seed pixel and all relevant point pixels, store the similarity calculation result S591, and form the similarity of the right image matching area S59 Set, the similarity calculation uses the following general formula:

$\mathrm{<span><span>\&rho;</span>\&rho;</span>}<span><span>=</span>=</span>\frac{\underset{\mathrm{<span><span>k</span>k</span>}<span><span>=</span>=</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>3</span>3</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}<span><span>[</span>[</span>{\mathrm{<span><span>g</span>g</span>}}_{\mathrm{<span><span>lk</span>lk</span>}}<span><span>-</span>-</span>{\stackrel{<span><span>\&OverBar;</span>\&OverBar;</span>}{\mathrm{<span><span>g</span>g</span>}}}_{\mathrm{<span><span>l</span>l</span>}}<span><span>]</span>]</span><span><span>[</span>[</span>{\mathrm{<span><span>g</span>g</span>}}_{\mathrm{<span><span>rk</span>rk</span>}}<span><span>-</span>-</span>{\stackrel{<span><span>\&OverBar;</span>\&OverBar;</span>}{\mathrm{<span><span>g</span>g</span>}}}_{\mathrm{<span><span>r</span>r</span>}}<span><span>]</span>]</span>}{\underset{\mathrm{<span><span>k</span>k</span>}<span><span>=</span>=</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>3</span>3</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}{<span><span>[</span>[</span>{\mathrm{<span><span>g</span>g</span>}}_{\mathrm{<span><span>lk</span>lk</span>}}<span><span>-</span>-</span>{\stackrel{<span><span>\&OverBar;</span>\&OverBar;</span>}{\mathrm{<span><span>g</span>g</span>}}}_{\mathrm{<span><span>l</span>l</span>}}<span><span>]</span>]</span>}^{\mathrm{<span><span>2</span>2</span>}}\underset{\mathrm{<span><span>k</span>k</span>}<span><span>=</span>=</span>\mathrm{<span><span>1</span>1</span>}}{\overset{\mathrm{<span><span>3</span>3</span>}}{\mathrm{<span><span>\&Sigma;</span>\&Sigma;</span>}}}{<span><span>[</span>[</span>{\mathrm{<span><span>g</span>g</span>}}_{\mathrm{<span><span>rk</span>rk</span>}}<span><span>-</span>-</span>{\stackrel{<span><span>\&OverBar;</span>\&OverBar;</span>}{\mathrm{<span><span>g</span>g</span>}}}_{\mathrm{<span><span>r</span>r</span>}}<span><span>]</span>]</span>}^{\mathrm{<span><span>2</span>2</span>}}}$

ρ为左图像某种子像素与右图像某相关点像素的相似度值，g_lk为某种子像素对应的红、绿、蓝灰度值，g_rk为某相关点像素对应的红、绿、蓝灰度值，k＝1、2、3，${\stackrel{\&OverBar;}{g}}_l=(g_{l1}+g_{l2}+g_{l3})/3,$${\stackrel{\&OverBar;}{g}}_r=(g_{r1}+g_{r2}+g_{r3})/3.$ρ is the similarity value between a sub-pixel of the left image and a pixel of a related point in the right image, g_lk is the gray value of red, green, and blue corresponding to a certain sub-pixel, and g_rk is the gray value of red, green, and blue corresponding to a pixel of a related point Gray value, k=1, 2, 3,${\stackrel{\&OverBar;}{g}}_l=(g_{l1}+g_{l2}+g_{l3})/3,$${\stackrel{\&OverBar;}{g}}_r=(g_{r1}+g_{r2}+g_{r3})/3.$

对相似度集合进行统计，采用求最大相似度的方法，如果最大相似度大于设定的阈值0.6，该相似度对应的右图像相关点像素为匹配点像素，否则，右图像匹配区域S59中无匹配点像素。遍历完左图像中的所有种子像素，依据相同的搜索策略，在右图像中找出所有种子像素的匹配点像素。The similarity set is counted, adopting the method of seeking the maximum similarity, if the maximum similarity is greater than the set threshold 0.6, the relevant point pixel of the right image corresponding to the similarity is a matching point pixel, otherwise, there is no matching point pixel in the right image matching area S59. Match point pixels. After traversing all the seed pixels in the left image, find out the matching point pixels of all the seed pixels in the right image according to the same search strategy.

4、匹配结果优化和纠错4. Matching result optimization and error correction

采用外极线约束条件和一致性约束原则对匹配结果进行优化和纠错，进一步排除匹配中的错匹配点像素和误匹配点像素。基于外极线约束条件，种子像素在右图像中的匹配点像素应该位于以种子像素为中心的纵向图像坐标可变的区域内，纵向图像坐标的浮动范围为±5像素，如果匹配点像素纵向图像坐标与种子像素纵向图像坐标的差超出该范围，即为错匹配点像素，否则为正确的匹配点像素。The matching results are optimized and corrected by using the epipolar line constraints and the principle of consistency constraints, and the wrong matching point pixels and wrong matching point pixels in the matching are further eliminated. Based on the constraints of the epipolar line, the matching point pixel of the seed pixel in the right image should be located in the region where the vertical image coordinates centered on the seed pixel are variable. The floating range of the vertical image coordinates is ±5 pixels. If the matching point pixel is vertical If the difference between the image coordinates and the longitudinal image coordinates of the seed pixel exceeds this range, it is a wrong matching point pixel, otherwise it is a correct matching point pixel.

基于一致性约束原则，左图像中存在先后位置的种子像素的匹配点像素也存在相同关系，如果匹配点像素的图像坐标不满足该关系，则为误匹配点像素。Based on the principle of consistency constraints, the same relationship exists between the matching point pixels of the seed pixels in the left image. If the image coordinates of the matching point pixels do not satisfy this relationship, it is a wrong matching point pixel.

5、物像反求5. Reverse object image

以左右图像中的种子像素和对应匹配点像素的图像坐标及视差为自变量，采用计算机双目立体视觉原理中弱视差显微立体视觉模型，计算出所有种子像素和匹配点像素对应的物空间相关点的空间坐标集合。若某种子像素的图像坐标为(w₁^l，w₂^l)，对应的匹配点像素坐标为(w₁^r，w₂^r)，视差定义为w₁^l-w₁^r。Taking the image coordinates and disparity of the seed pixels in the left and right images and the corresponding matching point pixels as independent variables, using the weak parallax microscopic stereo vision model in the computer binocular stereo vision principle, calculate the object space corresponding to all the seed pixels and matching point pixels A collection of spatial coordinates of the relevant points. If the image coordinates of a certain sub-pixel are (w₁^l , w₂^l ), the pixel coordinates of the corresponding matching point are (w₁^r , w₂^r ), and the parallax is defined as w₁^l -w₁^r .

对物像反求模型获得的物空间相关点的坐标集合进行处理，采用图形学中的三角剖分法对三维空间点造型，得到空间曲面。The coordinate set of the object space related points obtained by the object image inverse model is processed, and the three-dimensional space point is modeled by the triangulation method in graphics to obtain the space surface.

6、输出三维尺寸6. Output 3D size

选择三维重构面的被测区域，输出该区域的长度、宽度和高度方向的距离，该距离即为被测的三维尺寸。Select the measured area of the three-dimensional reconstruction surface, and output the distance in the length, width and height directions of the area, and the distance is the measured three-dimensional size.

对本领域的技术人员来说，很明显，本发明可以做出多种改进和变化，只要落入所附的权利要求书及其等同的范围内，本发明就涵盖本发明的这些改进和变化。It is obvious to those skilled in the art that various improvements and changes can be made in the present invention, and as long as they fall within the scope of the appended claims and their equivalents, the present invention covers these improvements and changes of the present invention.

Claims (8)

Translated fromChinese

1. 一种微结构三维尺寸立体图像快速检测方法，其特征在于：采用显微立体视觉系统采集立体图像对，并传送给计算机，计算机对立体图像对进行处理和匹配，具体处理步骤如下：1. a kind of microstructure three-dimensional size stereoscopic image fast detection method, it is characterized in that: adopt microscopic stereo vision system to gather stereoscopic image pair, and send to computer, computer is to stereoscopic image pair processing and matching, concrete processing steps are as follows:1)预处理图像；1) preprocessing image;对图像对中的左图像和右图像分别进行处理，得到左右预处理图像和左右目标图像；Process the left image and the right image in the image pair respectively to obtain the left and right preprocessed images and the left and right target images;2)提取图像中的特征点像素2) Extract the feature point pixels in the image提取左目标图像和右目标图像中的特征点像素，记录左图像中特征点像素的图像坐标和右图像中特征点像素的图像坐标；Extracting the feature point pixels in the left target image and the right target image, recording the image coordinates of the feature point pixels in the left image and the image coordinates of the feature point pixels in the right image;3)特征匹配3) Feature matching以左图像中的特征点像素为种子像素，在右图像中设置对应的搜索区域寻找其匹配点；Using the feature point pixel in the left image as the seed pixel, set the corresponding search area in the right image to find its matching point;4)匹配结果优化和纠错；4) Matching result optimization and error correction;5)物像反求；5) reverse object image;以左右图像中的种子像素和对应匹配点像素的图像坐标及视差为自变量，采用物像反求模型，计算出所有种子像素和匹配点像素对应的物空间相关点的空间坐标集合，采用图形学中的三角剖分法对三维空间点造型，得到空间曲面；Taking the image coordinates and disparity of the seed pixels in the left and right images and the corresponding matching point pixels as independent variables, the object image inverse model is used to calculate the set of space coordinates of all the seed pixels and the matching point pixels corresponding to the spatial coordinates of the object space. The triangulation method in science is used to model the three-dimensional space points to obtain the space surface;6)输出三维尺寸6) Output three-dimensional size选择三维重构面的被测区域，输出该区域的长度、宽度和高度方向的距离，该距离即为被测的三维尺寸。Select the measured area of the three-dimensional reconstruction surface, and output the distance in the length, width and height directions of the area, and the distance is the measured three-dimensional size.2. 根据权利要求1所述的一种微结构三维尺寸立体图像快速检测方法，其特征在于：步骤1中所述的预处理包括对左图像和右图像进行滤波、目标分割处理，区分开图像中的目标像素和背景像素，目标像素设置为黑色，背景像素设置为白色，得到左右预处理图像和左右目标图像。2. A kind of microstructure three-dimensional size stereoscopic image rapid detection method according to claim 1, it is characterized in that: the preprocessing described in step 1 comprises filtering left image and right image, object segmentation processing, distinguishing image The target pixels and background pixels in , the target pixels are set to black, and the background pixels are set to white, and the left and right preprocessed images and the left and right target images are obtained.3. 根据权利要求1所述的一种微结构三维尺寸立体图像快速检测方法，其特征在于：步骤2中所述的提取特征点像素的方法为求局部最大熵法或小波变换法或Sober算子法或灰度阈值法。3. a kind of microstructure three-dimensional size stereoscopic image rapid detection method according to claim 1, it is characterized in that: the method for extracting feature point pixel described in step 2 is to seek local maximum entropy method or wavelet transform method or Sober algorithm sub-method or gray threshold method.4. 根据权利要求1所述的一种微结构三维尺寸立体图像快速检测方法，其特征在于：步骤3中所述的特征匹配方法为：以左图像中的特征点像素为种子像素，在右图像中设置对应的搜索区域，遍历搜索区域中的所有相关点像素，计算种子像素和相关点像素之间的相似度，得到搜索区域对应的相似度集合，对相似度集合进行统计，获得搜索区域对应的区域相似度；把区域相似度与设定的相似度阈值比较，如果区域相似度大于相似度阈值，则当前种子像素的匹配点像素存在，否则匹配点像素不存在；遍历左图像中的所有种子像素，依据相同的搜索方法，在右图像中找出所有种子像素对应的匹配点像素。4. A kind of microstructure three-dimensional size stereoscopic image rapid detection method according to claim 1, it is characterized in that: the feature matching method described in step 3 is: take the feature point pixel in the left image as the seed pixel, and in the right Set the corresponding search area in the image, traverse all the relevant point pixels in the search area, calculate the similarity between the seed pixel and the relevant point pixel, obtain the similarity set corresponding to the search area, perform statistics on the similarity set, and obtain the search area Corresponding regional similarity; compare the regional similarity with the set similarity threshold, if the regional similarity is greater than the similarity threshold, the matching point pixel of the current seed pixel exists, otherwise the matching point pixel does not exist; traverse the left image For all seed pixels, according to the same search method, find out the matching point pixels corresponding to all seed pixels in the right image.5. 根据权利要求4所述的一种微结构三维尺寸立体图像快速检测方法，其特征在于：所述的对相似度集合进行统计的方法为最大相似度的方法。5. A kind of microstructure three-dimensional size stereoscopic image fast detection method according to claim 4, it is characterized in that: the described method for statistical similarity set is the method of maximum similarity.6. 根据权利要求4所述的一种微结构三维尺寸立体图像快速检测方法，其特征在于：所述的相似度阈值的取值区间为[0.6，0.9]。6. A method for rapid detection of three-dimensional images of microstructures according to claim 4, characterized in that: the value range of the similarity threshold is [0.6, 0.9].7. 根据权利要求1所述的一种微结构三维尺寸立体图像快速检测方法，其特征在于：所述的步骤4中采用外极线约束条件和一致性约束原则对匹配结果进行优化和纠错，排除匹配点像素集合中的错匹配点像素和误匹配点像素。7. A kind of microstructure three-dimensional stereoscopic image rapid detection method according to claim 1, it is characterized in that: in described step 4, adopt epipolar constraint condition and consistency constraint principle to optimize and correct the matching result , excluding the mismatching pixel and the mismatching pixel in the matching pixel set.8. 根据权利要求1所述的一种微结构三维尺寸立体图像快速检测方法，其特征在于：步骤5中所述的物像反求模型为小孔成像模型或投影模型或弱视差显微立体视觉模型。8. A method for fast detection of three-dimensional stereo images of microstructures according to claim 1, characterized in that: the object image inversion model described in step 5 is a small hole imaging model or a projection model or a weak parallax microscopic stereo visual model.

Images