基于机器视觉与光谱成像技术的玉米种子品质检测与分选

发布时间：2018-10-24 18:26

【摘要】：种子作为种植业最基本的生产资料,其品质优劣很大程度上决定了种植业的成败,直接关系到农民的增产、增收,影响到农村经济的稳健发展。种子质量检验评价是确保种子质量和种子资源管理的重要环节,对有效控制种子质量以提高产品产量和质量、贯彻优质优价政策、保证种子贮藏运输的安全及种子产业化进程的作用举足轻重。本研究以玉米种子为研究对象分别对其内部、外部品质进行检测,针对种子品质检测的难点和关键点,基于光谱成像技术、机器视觉技术,并结合图像处理技术、机器学习算法和光谱分析技术研究了玉米种子内外部品质(含水率、霉变、破损)的检测方法。基于上述研究的基础上开发的检测系统与算法为使用机器视觉与多光谱成像技术实现玉米种子品质快速检测奠定基础。本文主要研究内容和结论为:(1)研究了使用400-1000 nm高光谱图像进行玉米种子含水率的快速无损检测方法。首先,分别采集种子正、反两面400-1000 nm高光谱图像进行处理分析,提取质心区域光谱数据,采用竞争性自适应重加权变量选择算法(Competitive Adaptive Reweighted Sampling,CARS)筛选特征波段,建立对应的PLS含水率预测模型。其次,对比图像不同部位光谱曲线变化趋势,挑选2个特征波段(520 nm、560 nm)进行波段运算获取种子正、反面信息及质心位置。最后,依据正反面检测结果,自主选择对应的含水率预测模型对45个验证集样本进行含水率检测。结果表明:使用波段运算正、反面识别率分别为为97.8%、100%;正、反两面混合位置验证集相关系数分别为0.896、0.948、0.885,均方根误差RMSEV分别为0.823%、0.593%、0.858%;使用单面光谱信息建立含水率预测模型效果优于使用混合光谱信息建立的模型;使用反面胚乳区域光谱曲线建立含水率预测模型检测效果优于使用正面胚部位建立的含水率预测模型。(2)研究了使用1000-2500 nm高光谱图像进行玉米种子含水率的快速无损检测方法。分别采集种子正、反两面1000-2500 nm高光谱图像进行处理分析,建立正反面PLS含水率预测模型。使用4个特征波段(1104 nm、1304 nm、1454 nm、1751 nm)进行波段运算获取种子正、反面信息。结果表明:使用波段运算正、反面识别率分别为为97.8%、100%;正面、反面、混合位置验证集相关系数分别为0.969、0.946、0.947,均方根误差RMSEV分别为0.464%、0.616%、0.667%;使用单面光谱信息建立含水率预测模型效果优于使用混合光谱信息建立的模型;使用正面胚部区域光谱曲线建立含水率预测模型检测效果优于使用反面胚乳部位建立的含水率预测模型。(3)研究了使用机器视觉和图像处理技术进行玉米种子霉变及破损的快速检测与自动分选。使用经I-Relief算法筛选的有效特征参数训练分类器进行种子识别及检测,结果表明:提取全表面颜色特征参数训练分类器进行霉变玉米种子检测,使用朴素贝叶斯分类器进行霉变种子识别准确率达到98%;提取种子尖端颜色特征参数训练分类器进行霉变玉米种子检测,使用朴素贝叶斯分类器检测准确率达到99.3%;提取形状特征参数进行破损玉米种子检测,使用BP神经网络训练分类器准确率达到97.3%。
[Abstract]:As the most basic means of production in farming, the quality of the seeds determines the success or failure of the planting, directly affects the farmer's production increase, increases the income, and affects the steady development of the rural economy. Quality inspection and evaluation of seed is an important link to ensure seed quality and seed resource management. It is very important to control seed quality effectively to improve product yield and quality, implement high-quality and good-price policy, and ensure the safety of seed storage and transportation and the process of seed industrialization. Aiming at the difficulties and key points of seed quality detection, based on spectral imaging technology, machine vision technology and combined image processing technology, The detection methods of internal and external quality (moisture content, moldy and damage) of corn seeds were studied by machine learning algorithm and spectral analysis technology. The detection system and algorithm developed on the basis of the above-mentioned research laid the foundation for the rapid detection of maize seed quality using machine vision and multi-spectral imaging technology. The main contents and conclusions of this paper are as follows: (1) The rapid non-destructive testing method of corn seed moisture content using 400-1000 nm high-spectrum image is studied. First, processing and analyzing the high-spectrum images of 400-1000 nm on the positive and reverse sides of the seeds respectively, extracting the spectral data of the centroid region, screening the characteristic bands by using a Competitive Adaptive Weight-weighted Variable Selection Algorithm (CARS), and establishing a corresponding PLS water content prediction model. Secondly, two characteristic bands (520 nm, 560 nm) were selected for band operation to obtain seed positive, negative and centroid positions. and finally, selecting the corresponding water content prediction model to carry out water content detection on the 45 verification set samples according to the test result. The results show that the positive and negative recognition rates are 97. 8% and 100%, respectively. The correlation coefficients of the two-sided hybrid position verification set are 0. 896,0. 948, 0. 885, respectively. The root mean square error RMSEV is 0. 823%, 0. 593%, 0. 858%, respectively. The prediction model of moisture content is better than that of using mixed spectral information by using single-sided spectrum information, and the prediction model of moisture content is established by using the spectral curve of the reverse endosperm region. (2) The rapid non-destructive testing method of corn seed moisture content using 1000-2500 nm high-spectrum image was studied. and collecting the high-spectrum images of the seed positive and the reverse surfaces of 1000-2500nm respectively, and establishing a VPLS water content prediction model. Seed positive and reverse information was obtained using four characteristic bands (1104 nm, 1304 nm, 1454 nm, 1751 nm). The results showed that the positive and negative recognition rates were 97. 8% and 100%, respectively. The correlation coefficients of the positive and negative were 0. 969, 0. 946, 0. 947, and the root mean square error RMSEV was 0. 464%, 0. 616%, 0. 667%, respectively. The prediction model of moisture content is better than that of using mixed spectral information by using single-side spectrum information, and the model of moisture content prediction is established by using the regional spectral curve of frontal embryo section. (3) The rapid detection and automatic sorting of corn seed moldy and damage using machine vision and image processing technology were studied. Using the effective characteristic parameter training classifier screened by the I-Relieve algorithm for seed recognition and detection, the results show that the whole surface color characteristic parameter training classifier is used to detect the moldy corn seeds, and the accuracy of the moldy seed identification using the Naive Bayes classifier is 98%. The seed tip color characteristic parameter training classifier was extracted to detect the moldy corn seeds, and the accuracy of using Naive Bayes classifier was 99.3%, and the extraction shape characteristic parameters were used to detect the damage corn seeds. The accuracy of the classifier was 97.3% using the BP neural network training classifier.
【学位授予单位】：西北农林科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S513;TP391.41

【相似文献】