【作者】 王昱潭；

【导师】 李文彬； 阚江明；

【作者基本信息】 北京林业大学 ， 森林工程， 2014， 博士

【摘要】 灵武长枣是宁夏的重要经济林果,每年仅有20天左右的最佳采摘期。目前完全依靠人力架梯手工拣摘,采摘成本高、劳动力需求量大、强度高、且效率低。由于灵武长枣的经济价值高,种植面积还在逐年扩大。因此,对自动采摘技术的需求日趋强烈。自动化果蔬采摘机器人视觉系统首要解决的关键问题就是果实的定位与成熟度判别。本文主要研究在自然场景下,基于机器视觉的灵武长枣定位与成熟度判别方法,旨在为灵武长枣自动采摘机器人的研制奠定理论与技术基础。本文的研究内容及结果如下：1、在灵武长枣的灰度图像预处理中,提出了给定3×3滤波模板图像块中心阈值T,减少排序次数的中值滤波算法,提高了灰度图像处理速度；对于灵武长枣的彩色图像,结合HSI颜色空间各分量的特性,利用平均色度,提出了矢量结合标量的改进矢量中值滤波算法,对灵武长枣彩色图像进行降噪处理。2、根据灵武长枣的颜色特性,提出了两种在自然场景下的灵武长枣图像分割算法。(1)基于L*a*b*颜色空间各分量的特性,提出了基于L*a*b*颜色空间的给定阈值图像分割改进算法,红色分量阈值定为110。该分割算法能快速识别树上是否有成熟的长枣,算法的分割错误率为7.4%。(2)提出了基于色差融合的彩色图像分割算法,以颜色特征0.33R-0.5G+0.17B为基准,融合颜色特征R-G,分割红色目标区域；再以颜色特征0.122R+0.378G-0.5B为基准,融合颜色特征2R-G-B,分割绿色目标区域。该算法能很好的解决轻度的粘连、遮挡和光线不均等问题对图像分割的影响,并可以分别提取红色区域和绿色区域,分割成功率达到93.27%。3、经统计分析,得出了灵武长枣的横径与纵径的线性关系b1=1.64a+6.422；验证了灵武长枣呈近似椭球形,长枣体形与椭球体的拟合度超过90%；建立了灵武长枣外形的数学模型验证得出两个重要的结论：(1)灵武长枣的平面投影是近似椭圆；(2)在灵武长枣的平面投影中,近似椭圆具有短径不变的特性。根据这两个结论,推导了椭圆投影中心坐标；利用相机成像模型和基于最小二乘法的椭圆来拟合图像上的长枣、并确定长枣中心坐标,实现长枣在图像中的定位。4、建立了自然场景下灵武长枣生长成熟期的图像分类库；由实验测得的灵武长枣可溶性固形物含量和长枣表面颜色的红绿比,分析得到了二者与长枣生长成熟期的成熟度等级的关系；制定了基于机器视觉识别自然场景下生长成熟期的灵武长枣成熟度等级规则；建立了基于颜色的灵武长枣生长成熟期的成熟度颜色演化模型,提出了基于颜色演化模型的色调H与红色比相结合的灵武长枣成熟度等级识别算法,该算法的判别精度达到92.60%。本文的研究成果可解决自动采摘机器人视觉系统的关键问题,为林果采摘机器人系统奠定理论与技术基础。更多还原

【Abstract】 Lingwu long jujubes are important economical fruits in Ningxia with the best picking period of about only20days. Currently, their picking totally relies on humans, which is with large manpower demand, high labor intensity and low efficiency. The planting area is becoming larger year by year due to their high economic value. So there is an increasing demand for automatic picking techniques. The primary issue of establishing automatic picking robots based on machine vision is to solve the problem of localization and maturity recognition of targeting fruits. This dissertation studied how to automatically localize Lingwu long jujubes and recognize their maturity accurately and efficiently in natural scenes using machine vision, which should lay theoretical and technical foundations for making automatic picking robots of the jujubes.The main research contents and related results are as follows.1. For preprocessing jujubes’ gray-scale images, a novel median filter method combined the given threshold value T at the3×3filtering template block center was proposed to reduce sorting counts and accelerated the image processing as result; as for colored images, utilizing components of the HSI color space and the median chromaticity with combining the vector and scalar were implemented in the vector median filtering to annihilate noises of colored images of jujubes.2. According to the characteristics of color distribution of jujubes, two different segmentation approaches against the natural scenes were carried out.(1) The improved threshold segmentation:based on the L*a*b*color space, utilize the threshold value of red-green given as110to segment the mature jujubes quickly and the error rate reached7.4%.(2) The chromatism fusion segmentation:firstly, take the color feature0.33R-0.5G+0.17B as base to fuse the R-G feature in order to segment the red regions; secondly, take0.122.R+0.378G-0.5B to fuse the2R-G-B feature as to segment the green regions; in the end, combine red and green regions together to accomplish the segmentation of jujubes. This proposed method functions well against mild adhesion, occlusion and unbalanced light conditions, and its accuracy reached93.27%.3. Through statistic analysis, the linear mathematical model of the jujubes’ transverse and vertical diameters was established as b1,=1.64a+6.422, which verified that the shape of jujubes is more likely to be the ellipsoid and the degree of jujube-ellipsoid fitting achieved90%; the model of the jujubes’ shape was built as and two key conclusions were brought up:(1) the planar projection of the jujube should always be elliptical;(2) transverse diameter after the planar projection is invariant. Based on the two conclusions, the Lingwu long jujubes in images were localized in images using the camera pin-hole model and least square ellipse fitting.4. On the image recognition of jujubes in automatic picking, a catalog consisting of jujubes of different levels of maturity under natural scenes was established and classified accordingly as well; given soluble solid content of jujubes that was measured by experiments as well as facial colors with corresponded levels of maturity, a recognition hierarchy of maturity that is based on machine vision was constructed for jujubes growing in natural scenes and the evolutional model of jujubes’ maturity-color was also raised; based on the evolutional model, maturity recognition was implemented combining hue and red ratio and reached the accuracy of92.60%.The research results can solve the key problems of the vision system of the picking robots, which will provide both theoretical and technical foundations for the fruit-picking robot systems.更多还原