【作者】 姜莉；

【导师】 沈明霞；

【作者基本信息】 南京农业大学 ， 农业机械化工程， 2009， 硕士

【摘要】 农业收获机器人的正常工作有赖于对作业对象的正确识别,因而要实现对水果的采摘,前提是从果树中准确识别出水果。在本课题的研究中,提出了基于人类视觉特性的自然光照条件下水果识别方法,并用于树上成熟石榴的检测试验。本文主要研究内容和方法如下：(1)首先,基于人类视觉系统对颜色感知特性,在HSI空间对石榴采摘图像进行特征研究；通过试验分析光照对饱和度和色调的影响,并基于人类视觉特性提出解决光照问题设想。结合颜色恒常性理论,提出了一种将全局特征和局部特征相联合的带颜色恢复因子的多尺度Retinex(MSRCR)算法。实验表明该算法实现了动态范围压缩、色彩恒常、颜色高保真等特点的自动图像增强。(2)针对果实图像的灰度和颜色特点,利用色调直方图呈明显双峰状,采用OSTU自动阈值分割,有效去除大部分背景。对二值图像首先采用形态学滤波去除残留小面积噪声；八邻域标记并用种子填充算法进行区域填充；然后统计各标记区域面积,根据区域面积和外接矩形长宽比作阈值,将不符合参考值的区域作为伪目标去除,从而使分割效果得到进一步改善；采用Sobel算子进行轮廓提取；利用优化后的广义Hough变换,有效地恢复果形并准确提取果实半径及形心。(3)利用图像的不变矩,并结合形态、灰度和变换域信息,通过构造综合特征来进行图像匹配,并将其应用于目标检测。即：在图像预处理时,利用SUSAN算子检测图像目标的边缘和角点,来实现初定位与分割；引入不变矩理论,通过实验证明面矩和线矩都具有平移、比例放缩和旋转不变性,基于线矩的特征矢量计算在时间上明显优于基于面矩的计算,有利于提高图像处理的实时性。然后基于线矩提取局部显著性特征,并结合面积周长比、长宽比、占空比、灰度信息等综合特征进行图像匹配,实现了成熟石榴识别,并给出具体实现过程及结论。更多还原

【Abstract】 The normal work of agricultural harvesting robot depends on correct recognition of manipulating object, the fruit detected accurately in a tree is the prerequisite of the successful harvest. The recognition method of fruits in natural lighting conditions based on human visual characteristics is proposed.In this research, the detection experiment of mature pomegranate was conducted in nature. The main research contents and methods are as follows:(1)Firstly, characteristics of picking pomegranate image were studied based on color perceptual characteristics of human visual system in HSI space model. The illumination influences about saturation and color were analyzed by experiments, at the same time, solving illumination theoretical assumption was proposed based on human visual system. Secondly, a multi-scale Retinex algorithm with color healing factor(MSRCR) composed global and local characteristics was studied. Combined with color constancy theory, the experiments showed that the algorithm has realized the automatic image enhancement with different characteristics, such as dynamic range compression, color constancy and color high fidelity.(2) According to gray and color characteristics in fruit image, the image histogram of hue showed a clear bimodal shape obviously, so the OSTU automatic threshold segmentation was adopted in order to remove most of image background. Firstly, morphological filtering was used to remove residual small area noise in binary images. Secondly,labeled by eight neighborhoods,optimized seed filling algorithm was used to fill the holes. Thirdly,every labeled area was calculated out, set the threshold according to the area and the length to width ratio of minimum enclosing rectangle, and the other areas were wiped out. The edge was detected by Sobel algorithm.At last,centroids and radius were extracted by the optimized Hough transform, and the shape of fruits was recovered.(3) The comprehensive characteristics can complete image matching based on invariant moments of combination with configuration and intensity and transform domain, and applied to the target recognition. That is, Edge and corner point detection used SUSAN operator at low level image processing, in order to realize the preliminary localization and segmentation; while the invariant moments theory were introduced, invariant moments characteristics of area moments and line moments theory were proved by mass data tests, when translated or constrained or rotated.And the running average time of line invariant moments is more superior than area invariant moments, it is helpful to improve real time of image processing. Then the local significant characteristics extracted by line moment invariants, and image matching combines with the comprehensive characteristics, for instance, area to perimeter ratio, length to width ratio, duty ratio, intensity information etc. The detection of mature pomegranate was implemented, also the realization process and conclusions of target recognition were presented in this paper.更多还原