【作者】 谢忠红；

【导师】 姬长英；

【作者基本信息】 南京农业大学 ， 农业电气化与自动化， 2013， 博士

【摘要】 中国是一个水果生产大国,利用机器视觉技术实现自动收获能够解决我国劳动力短缺问题并大幅提高生产力。但由于果树的生长环境的非结构化,采摘机器人的视觉系统获取的图像是一幅包含有天空、树枝、树叶、土壤和果实的复杂图像,且树上果实的生长形态多种多样,果园的光线变化无常,所有这些因素决定了建立一个实用摘机器人视觉系统是一个复杂的系统工程。本文以在自然光条件下拍摄的果树图像为研究对象,结合图像处理、计算机视觉以及人工智能技术对果实采摘机器人视觉系统中图像处理关键算法进行了深入的研究,主要工作和研究成果如下：1)基于改进k-mean均值聚类算法的彩色果实图像分割首先对颜色空间进行了深入研究,在RGB, HIS, YCgCr, YCbCr和CLE5种颜色空间中选择21个颜色特征,利用BP神经网络系统找出识别率最高、误分率最低的颜色特征组(H,Cr(YCgCr), Cr(YCbCr), R-G,2R-G,Cb-Cr),并将该特征组作为果实图像分割的向量；然后利用变异系数赋权法对k-means聚类算法进行改进,并对识别样本集中的果实图像进行分割试验。结果表明,本文提出的算法能够对自然光照条件下成熟桃子图像进行有效分割,且对逆光拍摄的桃子图像分割效果也很好,算法识别正确率大于89.5%。该方法结合了多种颜色空间的优势,克服了传统欧氏距离在聚类算法中的缺陷,提高了图像分割的效率和识别精度。2)结合颜色和纹理特征两步法实现果实和背景精准分割以自然条件下采集的果实图像为对象,研究利用颜色和纹理特征进行果实识别的方法,提出了一种基于纹理和颜色两步图像分割方法。首先选择16×16,8×8两种窗口对图像进行分块处理,计算基于灰度共生矩阵(0°,45°,90°,13504个方向)的能量(en)和对比度(con)特征；同样对这两种窗口图像进行一次Harr小波分解,获取4种纹理特征：水平细节图像对应的能量(enChl)和对比度(conChl),垂直细节图像对应的能量(enCvl)和对比度(conCvl)。利用BP神经网络选择最优纹理分割特征：16×16窗口的对比度(con)和垂直细节的对比度(conCvl),利用训练好的神经网络对图像进行一次分割,粗略确定果实位置；然后颜色特征H和R-G对图像进行二次确认分割。研究结果表明,不管是顺光图像还是逆光图像,两步法均能达到很好的分割效果。3)改进的圆形Hough变换检测并定位独立生长状态的类圆果实首先为了能够快速准确地计算出相互分离且无遮挡类圆果实的形心坐标和半径,提出来了一种基于改进圆形随机Hough变换的快速类圆果实目标检测方法。在实现背景分离后获取单像素果实轮廓,并按步长获取果实的边缘特征点；然后,根据边缘特征点的平均切线方向对特征点进行分组,并以此为依据对圆形RHT算法进行改进；最后利用改进后的圆形RHT算法计算出类圆果实的形心坐标和半径。研究结果表明,该方法能够快速准确地对生长状态为相互分离的类圆果实进行检测,对部分被遮挡的类圆果实识别效果也较好。4)基于凹点搜索的重叠果实的快速定位和检测为了能够快速准确地计算出生长状态为靠拢或重叠的成熟类圆果实的形心坐标和半径,提出了一种基于凹点搜索的快速定位和检测重叠果实目标的方法。采用链码跟踪法获得单像素果实轮廓,并按步长获取边缘特征点；然后利用N点方向编码差获取边缘凹点,确定凹点群,根据阈值确定分割凹点；最后利用改进的Hough变换计算出多个重叠桃子的形心坐标和半径。研究结果表明,该方法能够快速准确地实现多果重叠情况下果实目标的检测。5)基于相位编组的树枝障碍物的检测方法为了能够快速准确地计算出果树枝干的空间位置,提出了一种基于梯度相位编组的Hough变换树枝检测新算法。利用改进的平方梯度法计算边缘点的梯度相位角,找出梯度相位直方图中多个阈值大于T的峰值。然后将具有相近梯度相位角的边缘点归为一组。最后对每组中的边缘点采用改进的两点表决Hough变换算法找出对应的直线参数。利用梯度相位角进一步验证参数的正确性。研究结果表明,本文提出的梯度相位编组直线检测方法具有速度快,检测误差小和鲁棒性强的特点,该方法能够快速准确地实现果树树枝的定位和检测,对部分被遮挡的树枝识别效果也较好。6)研究了独立生长的果实的生长姿态的识别方法为了避免机械手在收获过程中因缺少果实生长姿态信息造成对果实和枝干的损伤,提出了一种苹果生长姿态估算方法。首先以无遮挡相互分离的树上苹果为研究对象,根据果实轮廓的几何特征提出了四种判定果实生长姿态的方法①二阶中心矩法——由2阶中心矩计算出的惯性主轴方向就是苹果的果轴方向；②最短距离法——利用最短距离原理找出质心和果梗中心点,由这两点定果轴方向；③最小斜率方差法——斜率方差最小的边缘轮廓段的中点,被视为果梗中心点。④三点一线法—果梗中心点,质心和花萼遗迹中心点三点可以确定果轴方向。最后,分析了已有的四种方法的优缺点,并提出了综合四种方法进行决策融合。研究结果表明,基于四种方法的决策融合识别苹果的正确率高于单独使用任何一种方法可达到了90%,且识别速度降幅在允许范围内。更多还原

【Abstract】 China is a large country where a lot of fruits grow in.In order to alleviate shortage of the labour force and raise labor productivity, machine vision technology is applied in fruit atomatic picking system. Generally, for the fuit trees grow in the non-structural environment, the image captured by computer vision system of picking robot is composed of sky image, branches image, leaves image, soil image, fruits image and et al. The varieties of growth morphology of fruits and the changeable sunlight determin that buiding a vision system is a complex system engineering.Regarding images of fuit trees taken in natural scene as the research object, with the help of technology of image processing,computer vision and artificial intelligence, the key technology of vision system of fruit picking robot was studied in this paper, the main work and reserch results were as following:l)Color Image Segmentation based on improved K-means algorithmFirstly, after making an intensive study on the color spaces,21color characters were selected from RGB,HIS,YcgCr,YcbCr and CIE five color spaces. The combination of color characters (H,Cr(YCgCr), Cr(YCbCr), R-G,2R-G,Cb-Cr) group which had maximum recognition rate and minimum recognition error rate were found based on BP neural network, and the combination of color characters was used as the feature vector to segment the fruit image; Secondly, images in recognition sample space were segmented by using the improved k-means algorithm based on variance coefficient weighting method. The overall results showed that the proposed method could segment the ripe peach image under natural sun light validly, it could also segment the images taken opposite the sun, the recognition accurate rate could obtain89.5%. The method proposed in this thesis which combined the advantage of several color spaces,could raise the image segmentation efficiency and identification accuracy.It could also overcome the defects of trandional euclid distance in clustering algorithm.2)Two-step method based on texture and color characters to separate the fruits from the background acculatelyRegarding the fruit images captured in natural scene as the research subject, the fruit recognition methods basedon texture and color charcter were studied, and two-step method based on texture and color characters to separate color image was put forward. Firstly, image was divided into blocks whose size were16×16and8×8, energy and contrast based on Gray-level Co-occurrence Matrix((0°,45°,90°,135°) were calculated in the two kinds of block. Four kinds of texture character:Enger of horizontal detail image, contrast of horizontal detail image, enger of vertical detail image and contrast of vertical detail image were obtained after the image was first decomposed with Haar wavelet in these two kinds of block. After the best texture characters:con and conCvl of16×16block were selected by BP network, the position of fruit in image was located approximately by using the trained BP neural network to segment the image. Then the image was segmented again by use of the color characters:H and R-G. The results showed that two-step method could get good segmentation results for images opposite the sun or front the sun.3) Location and detection for single quasi-circular fruits based on improved hough transformIn order to calculate accurate centric coordinates and radius of quasi-circular fruit rapidly, a kind of detection method for quasi-circular fruits based on improved circular randomized Hough transform was proposed. After the object was segmented from background with2R-G, the thinning algorithm was used to extract one-pixel fruit contour, from which the edge character points were abstracted. Then the edge character points were grouped according to their average tangent directions, with which the circular RHT algorithm was improved. Last, the centric coordinates and radius of quasi-circular fruits were calculated with the optimized circular RHT algorithm. The overall results showed that the proposed method could detect the quasi-circular fruits rapidly and accurately, it could also recover the shape of part-covered fruit satisfactorily.4)Location and detection for overlapped fruits based on searching concave spotsIn order to calculate centric coordinates and radius of quasi-circular ripe fruits whose growing state was approached or overlapped, a kind of fast location and defection method for fruits object based on searching concave spots was proposed. After the object was segmented from background with hue according to statistical law, the freeman chain code algorithm was used to extract one-pixel fruit contour, from which the edge character points were abstracted. Then the edge concave spots which were found by direction coding difference of N points, were divided into several concave spot groups, and the segmentation concave spots were located by the threshhold. Last, the centric coordinates and radius of several overlapped peaches were calculated based on optimized circular Hough transform. The research results show that the proposed method can detect the overlapped ripe fruits accurately and rapidly.5)Location and detection for branches based on gradient phase grouping A new method of Hough transform based on gradients phase grouping for branches detection was proposed to locate the position of branches of fruit trees accurately and rapidly. After the gradient phase of edge points were calculated by using the squared gradient method, the histogram of gradients’s direction were calculated. From the histogram, several peak values were found by threshold T. Then edge points were grouped according to their gradent phases, and points in each group almost had the same gradient’s direction. Last, an improved two-points Hough transform was applied to edge points in each group to calculate the parameters of lines, and the gradent’s direction of every group was used to affirm the correctness of the parameters. The research results showed that the proposed had the merits of high speed, low error and strong robustness, and could locate the branches of fruit trees accurately and fastly,it could also detect part-covered branches satisfactorily.6) Study on methods to estimate the growing attitude of single appleIn order to avoid damaging apples and branches caused by manipulator during the picking operation, for the absence of attitude information, an apple’s attitude estimation method was put forward. After the apple object was segmented from background with the two-step algorithm based on the characters of color and texture, the freeman chain code algorithm was used to extract one-pixel fruit contour. Then least distance method, least slope variance method and three collinear points method were given,and the recognition rates of three methods were compared. Lastly, for the purpose of improving recognition rate, decision method based on fusion of four methods was proposed. The research results showed that the recognition rates by use of four methods were higher than using any of methods seperatly, and the recognition rates could reach90%.更多还原