【作者】 李江波；

【导师】 应义斌； 饶秀勤；

【作者基本信息】 浙江大学 ， 农业电气化与自动化， 2012， 博士

【摘要】 水果表面缺陷是决定水果价格最有力的因素之一,因为外部缺陷是对水果品质最直接的反映。国家标准对水果表面缺陷数量和面积大小有严格的规定。与水果的大小、颜色、形状等外部质量指标相比,水果表面缺陷的快速识别一直是水果分级中最难,耗时最多、研究人员最感兴趣的研究内容。多年来,研究人员做了大量的工作。本研究以脐橙为研究对象,利用RGB成像技术、可见近红外高光谱成像技术及荧光高光谱成像技术,详细地探讨了脐橙表面11类型常见缺陷(包括蓟马果、溃疡果、裂伤果、炭疽病果、日灼果、药伤果、风伤果、虫伤果、介壳虫果、异色条纹果和腐烂果)检测理论以及两种较为重要的缺陷溃疡病和腐烂果的识别理论。解决了目前脐橙表面缺陷检测的部分难题,所开发的脐橙表面缺陷检测算法对下一步自动、快速、在线脐橙缺陷分级装备研发奠定了重要的基础。主要研究内容和研究结论如下：(1)提出了掩模法去背景的背景分割理论。利用构建的二值化掩模模板对脐橙图像进行去背景,通过对静态图像和在线图像背景去除结果表明,背景去除率达到100%的同时,脐橙表面信息可以较好地保留。这为脐橙表面缺陷进一步有效提取奠定了基础。(2)提出了一种新颖的脐橙表面亮度不均变换的照度-反射模型及单阈值快速水果缺陷分割算法。基于此亮度变化模型,正常水果表面区域被提升为高亮区域,而水果表面的缺陷区域依然保持低灰度,这一变化克服了由于类球形水果表面亮度分布不均导致缺陷分割精度低的难题,这也为单阈值脐橙表面缺陷快速分割提供了可能。试验表明,与边缘亮度补偿算法相比,该理论可以对脐橙表面整体亮度进行变换,并且该亮度变换理论比B样条曲面拟合理论处理一幅图像速度超过30倍,基于此算法及单阈值分割理论对风伤果、蓟马果、介壳虫果、裂伤果、炭疽病果、日灼果、溃疡果、异色条纹果、虫伤果、正常果和药伤果等11类型共计6345个感兴趣区域进行分割,获得了93.8%分割精度。(3)通过对不同类型缺陷RGB图像不同灰度值统计后获得区分脐橙果梗与缺陷的算法。基于该公式和提出的大区域及长区域去除算法BER可以获得100%的果梗识别率且不会受到其它缺陷类型的影响。由于该算法仅仅涉及两次减法、一次乘法及一次除法,避免了复杂的模式识别理论,所以具有一定潜力应用于脐橙缺陷在线检测。(4)本研究开发了脐橙表面缺陷检测联立算法,此算法主要有四个模块构成,即背景分割模块、亮度不均变化模块、果梗识别模块和果脐识别模块。应用此算法针对11类1320幅样本图像识别结果表明,99.1%的缺陷果是被正确识别为缺陷果,98.3%的正常果被正确识别为正常果。另外,通过调节不同的阈值可以满足不同的用户需求。(5)研究发现可见近红外光谱区域的6个特征波长(630、691、769、786、810和875nm)或者3个特征波长(691、769和875nm)能被用于构建多光谱脐橙表面缺陷检测系统。针对9种带有不同表皮的橙子样本,利用研究中所开发的双波段比和主成分分析相结合的缺陷果检测算法,获得了最好93.7%的橙子正确识别率,并且假阳性率为0。与模式识别算法相比,双波段比图像R875/691能有效地区分果梗和缺陷区域,因为模式识别理论增加了算法的复杂度。(6)搭建了荧光高光谱成像系统,采用该系统对脐橙早期腐烂缺陷进行了研究。利用最佳指数OIF理论获得了识别腐烂果的最优波段,即498.6nm和591.4nm,该方法克服了高光谱图像数据量大和相邻波段之间的强相关性等不足,实现了高维数据的降维,快速确定了特征波长。基于特征波长比图像及双阈值分割算法,获得了整体100%的腐烂果识别率。同时,该双阈值理论较好地避免了梗伤缺陷的荧光效应对腐烂果检测的影响,从而降低了系统及算法成本。(7)搭建了可见近红外高光谱反射成像系统,采用该系统对脐橙溃疡果识别进行了研究。研究获得了用于溃疡识别的7个关键波段(630、687、765、788、815、833和883nm)。基于此7个波段的第三主成分图像和2个波段的波段比图像Q687/630开发了多光谱溃疡检测算法。带有11种类型共计275个独立样本用于检测算法的可行性,获得整体98.2%溃疡果识别率。本研究也发现,单独的双波比Q687/630理论对于区分溃疡果和除炭疽病及日灼伤外的其它类型果可以获得97.8%的识别率。以上的研究成果为我国研发基于机器视觉技术的脐橙表面缺陷在线、快速检测分级装备奠定了重要的基础。更多还原

【Abstract】 The presence of surface defects is one of the most influential factors in the price of fruit, since most consumer associates quality with a good appearance and the total absence of external defects. Chinese national standard has a strict rule of number and area of fruit surface defect. Compare with other external quality indexes, such as size, color and shape, fast detecting defect on fruit always is a most difficult, time-consuming and interesting task for researcher. Researchers have done a lot of work for years.Navel oranges are used as samples in this study. Detection methods on eleven types of common surface defects (i.e., thrips scarring, canker spot, dehiscent fruit, anthracnose, copper burn, phytotoxicity, wind scarring, insect damage, scale infestation, heterochromatic stripe and rottenness) and identification methods on two very seriously diseases (canker spot and rottenness) were detailedly studied by using RGB imaging, Vis-NIR hyperspectral imaging and fluorescence imaging. Some problems in detecting navel orange surface defects detection were solved in this study. The developed detection algorithm offered help for developing an automatic, fast and on-line navel orange surface defect detection system. Main contents and results were listed as follows:(1) A method for removing fruit image background by using mask template was introduced. Binary mask template was firstly built. Then, the mask was used to perform background segmentation of static and on-line fruit images, respectively. The results showed that background segmentation rate was 100%, and the information on fruit surface can be kept perfectly. This laid the foundation for further effectively segmenting navel orange surface defects.(2) A novel illumination-reflectance model for correcting non-uniform light on navel oranges and segmentation algorithm based on single-threshold value for extracting defects was proposed. Using this light transform model, the intensities of normal areas on navel orange surface were changed into high values. However, defective areas were still kept low intensities. This light transform overcame low defects segmentation accuracy that was caused by non-uniform light distribution on spherical fruit surface. And, it was also useful for fast segmenting navel orange surface defects based on single-threshold value method. The study results showed that light transform model could correct the whole non-uniform light distribution on navel orange surface compared with edge light compensation algorithm, and its speed in processing an image was more than 30 times comparing with B spline curve fitting method. In order to evaluate performance of this algorithm,6345 regions of interest from 11 types of samples were firstly marked. Then, these marked regions were segmented using the developed algorithm. The results showed that 93.8% segmentation accuracy was achieved.(3) An algorithm to differentiate stem-end from different types of defects was proposed based on calculating intensities of different types of defects from R, G and B component images. The study results showed that 100% stem-end identification rate was obtained using proposed algorithm and developed a big area and elongated region removal algorithm (BER). The stem-end identification method is potential to be used in an on-line neval orange defect sorter since it only includes two subtraction operations, one multiplication operation and one division operation avoiding complex pattern recognition methods.(4) A combination algorithm for deftecting defects on neval orange surface was proposed in this study. The algorithm includes four modules such as background segmentation, light non-uniform transform, stem-end identification and navel identification. The results from 1320 sample images includes 11 types of defects showed 99.1% of fruit with defects and 98.3% of normal fruit were correctly identified. In addition, different requirements of users can be met by adjusting threshold values.(5) Study found that six characteristic wavelengths (630,691,769,786,810 and 875nm) and, alternatively, three wavelengths (691,769 and 875nm) in the visible and near-infrared spectral range could be potentially implemented in multispectral imaging systems for detecting orange peel defects. The defect detection algorithm combining two-band ratio with PCA achieved 93.7% identification rate for orange surface defects and no false positives. It should also be pointed out that simple two-band ratio (R875 691) algorithm could be more effective to identify stem-ends from skin defects compared to pattern recognition algorithms, which increased the computational complexity.(6) Fluorescence hyperspectral imaging system was developed to detect early rottenness on oranges. The optimal band (498.6nm and 591.4nm) for identifying rottenness defects was obtained by using Optimum Index Factor (OIF) method which can overcome some disadvantages caused by large data quantity and strong relativity between adjacent bands from hyperspectral images. Based on ratio images and a segmentation algorithm with double thresholds,100% fruit with rotten area was identified. In addition, the algorithm with double thresholds could also effectively avoid the influence from stem-end fluorescence effect. Therefore, the cost of system and algorithm was decreased.(7) Vis-NIR hyperspectral reflectance imaging system was developed to identify oranges with canker spots from those with other types of peel. Seven important wavelengths (630,687, 765,788,815,833 and 883nm) were obtained in this study. The third principal component image based on obtained seven bands and ratio image Q687/630 based on obtained two bands were used to develop multispectral canker identification algorithm.275 independent samples with 11 types of peel defects were used to estimate the feasibility of algorithm. An overall 98.2% canker identification rate was obtained. In addition, in this study, it should be also noticed that two-band ratio images give better recognition results (97.8%) of discriminating canker from normal and other diseased skin conditions aside from anthracnose and copper burn.The above work provided an important foundation for developing on-line and fast detection equipment of navel orange surface defects using machine vision technology in China.更多还原