【作者】 冯斌；

【导师】 汪懋华；

【作者基本信息】 中国农业大学 ， 农业电气化与自动化， 2002， 博士

【摘要】 利用计算机视觉技术进行水果品质的在线检测与分选技术研究，对提高果品市场竞争力与产品增值效益具有重要应用前景。特别是在我国加入WTO世界贸易组织之后，这一需求显得更为迫切。本文就是在这样的背景下，研究了水果在线检测与品质分选的方法和技术，目的在于解决动态条件下，图象质量差，信息量大，实时处理能力低，检测精度低等问题。主要研究内容如下： (1)在分析现有微分边缘检测算子的基础上，根据水果图象在检测时只需要检测目标外边缘的特点，提出了两种新的边缘检测算法：灰度邻域法和模板分析法。这两种方法检测的图象面积仅约传统方法的1／2，因此检测速度约是传统方法的2倍。由于在搜索过程中采用圆周上等角度方向进行，使检测出的边缘点有序。且边缘点清晰、连续，无需进一步细化和序列化处理，提高了系统处理速度。 (2)根据试验图象背景的均匀特性，提出了图象的快速定位和标记方法。对160×140大小的图象以10×10网点处理，仅需处理224个象素点后，就可以通过质点法计算出目标物体的参考形心和参考平均半径，有效地减少了后续处理的图象面积，提高了处理速度。 (3)对于运动造成的图象模糊问题，传统方法是根据运动成像模型分析得出的差分算法进行恢复。但是该方法计算量大、实时性差，且恢复过程属近似计算过程。本文在运动成像模型的基础上，结合图象的特点，以图象象素分析的方法进行恢复。试验结果表明恢复效果好，且处理速度快。 (4)对果径的检测提出了新的轴向检测算法和果径检测方法。克服了动态检测过程中传统方法很难确定果梗而造成果径检测的误差。该方法检测的方向与国标要求的方向一致，通过实际测量，计算机检测结果与人工检测结果有良好的相关性。本中提出的水果果形模型使形状的描述从定性提高到定量分析的水平，为果形分级提供了依据，使形状分级过程简化而且有效。 (5)提出以各色度域的分形维数为颜色特征值，取4个域的分形特征值作为输入模式，以人工神经网络进行颜色分级。由于这些特征值在考察各色度值累计特性的同时，考察了各色度的空间分布特性，因此，使颜色分级过程更符合实际情况。 (6)提出以待测图象的反射特性、平均半径为参数的标准球体灰度模型。以该模型的灰度值与待测图象作差进行缺陷分割，仅用单个阈值使不同灰度级的缺陷一次分割成功。该方法计算量小，操作速度快，同时不会在边缘产生接缝问题。 (7)提出了以水果空间结构特点识别缺陷与果梗花萼的方法。该方法取可疑缺陷区边缘上、下、左、右4个方向的灰度剖面线平均后作为特征剖面线，通过傅氏变换，再以低频项系数进行傅氏反变换重建，得剖面的总体形状，用该形状来识别缺陷。 (8)建立了以分级为目的的软硬件系统。硬件系统可完成水果的传输和动态捕获图象的功能。软件系统包括水果大小、形状、颜色以及缺陷的分级功能。更多还原

【Abstract】 Study on the real time fruit quality detection by computer vision is an attractive and prospective R & D subject for improving marketing competition and post harvesting value-added processing technology of fruit products. As China entering WTO,it becomes more and more urgent. The objectives of this research are contributed to develop method and technology for fruit on line detection by computer vision. It aims at solving the problems,such as fast processing the large amount of image information,improving system performance for real time dynamic image capture and processing capability,increasing precision of detection and on line grading system establishment,etc. The results of study are briefly summarized as follows:1. Based on analysis of current differential edge detection arithmetic operators and the requirement to detect only outer edge of objects in fruit detection,two new methods - gray adjacent area and template analysis were introduced for the solution. The image area detected with the new methods is only equal to half of traditional way,but the processing speed can be doubled. Because the search was carried out along equal angle on a circle,the edge point detected can be clear ordered and keeping continuity,so the farther thinning and serial processing were not needed and the processing speed of system was much improved.2. Based on the uniformity of image background,a method of quick image orientation and marking was put forward. The 10 x 10 grids can be used to deal with the image of 160 x 140. Only after processing 224 pixels,the reference figure center and average radius of object can be calculated by particle method. It is very effective to reduce processing area and to improve processing speed.3. With the problem of blurred image caused by object motion,traditional difference algorithm based on analyzing the model of moving image was generally adopted to recover the blurred images. But its calculation work is too much and the capability of real time processing is bad. The resume process actually is as approximate calculation. A new method in this paper was presented to resume the original image based on the pixel analysis. The method of motion imaging and combining with the characteristic of image could be resumed by pixels decomposing. The test showed good features and processing speed is quite fast.4. In order to measure the size of fruit,a new way for the measurement of axis direction and width was presented in this paper. In the traditional way,the stem-end can not be ascertained,which would cause measurement error. The new method may overcome this disadvantage. The measurement direction of new method was consistent with national standards. By the measurement experiment,the results of computer detection has better than the human operation. The presented model of fruit shape can improve the shape description both in qualitative and quantitative analysis. The model is used as basis for shape classification. It may simplify the classification process and make the process more effective.5. The fractal dimension of every hue area was considered as color feature value. The fractal featurevalue of four hue area were used as input mode. The color was graded by artificial neural networks. Because of considering accumulative character and space distributing character of each hue at the same time,these feature values can make the color classification more close to reality.6. A normal sphere hue model based on reflect character and average radius of image was introduced. It is used for fruit defects segmentation through comparison between the gray values of model and the detected object image. The defects were divided by the gray value of the image. Only one threshold is needed for a successful segmentation of the defects flaw of every gray level. In this method,less calculation is required and the processing speed is faster. There was no any juncture on edge.7. A method based on the sphere frame character was used to identify the surface defects an更多还原