鸡蛋多品质高通量在线快速无损检测研究

发布时间：2018-06-25 19:53

本文选题：鸡蛋 + 无损检测　；参考：《华中农业大学》2017年博士论文

【摘要】：鸡蛋品质检测是鸡蛋商品化处理中的关键环节,对提高鸡蛋的经济价值和改善人们生活品质有着重要意义,尤其是高通量在线检测,对提升我国的鸡蛋加工生产自动化水平和鸡蛋产业发展具有积极作用。为了实现鸡蛋品质的高通量在线快速检测,本课题结合鸡蛋加工实际生产需求,以鸡蛋的新鲜度、散黄、尺寸形状、破损等多个品质为研究重点,利用光谱分析和机器视觉技术对鸡蛋的多个品质进行检测。主要研究内容和结论如下:1)鸡蛋新鲜度的光纤光谱快速定量检测。利用自行搭建的光纤光谱检测装置采集鸡蛋透射光谱信息,结合Savitzky-Golay卷积平滑滤波、多元散射校正、标准正态变换、一阶微分及二阶微分5种预处理方法,分别建立了偏最小二乘回归PLSR和支持向量回归SVR模型,比较不同模型精度,发现一阶微分处理的SVR模型预测效果较好,且SVR模型在总体上优于PLSR,表明SVR能够较好地提取鸡蛋新鲜度与光谱信息之间隐含的非线性关系;为了简化定量模型来达到快速检测鸡蛋新鲜度,使用线性降维主成分分析法PCA和流形学习中的非线性降维局部线性嵌入LLE分别对一阶微分后的光谱数据再处理,比较了两种降维后的模型结果说明LLE更好地提取了光谱有效信息,提高了模型精度,降维效果比PCA更加明显。LLE-SVR模型中训练集和预测集相关系数和均方根误差分别为0.922、7.21和0.911、8.80,交叉验证均方根误差比PCA-SVR下降了0.79。研究结果表明LLE-SVR模型可以用于光纤光谱快速定量检测鸡蛋新鲜度,为今后鸡蛋新鲜度的高通量在线检测作了理论研究。2)散黄蛋的光纤光谱快速在线识别。利用光纤光谱技术在5000枚蛋/h单通道的传输装置上动态采集鸡蛋透射光谱数据,比较了连续投影算法SPA和竞争性自适应重加权算法CARS分别对不同光谱预处理数据的波长优选情况,发现SPA选取的特征波长个数总体低于CARS,然后结合所选的特征波长采用偏最小二乘判别PLS-DA、分类回归树算法CART、K近邻分类算法KNN和簇类独立软模式算法SIMCA四种分类方法建立多个分类器,根据变量个数和识别率比较分类器性能,优选出5个分类器,最后通过比较每个分类器对单枚鸡蛋的检测时间,确定SNV-SPA-PLS-DA模型适用于在线识别散黄蛋,其特征变量只有3个,单个鸡蛋检测时间为55.733ms,预测准确率达到97.14%,为散黄蛋高通量在线光谱识别提供技术方法。3)鸡蛋尺寸形状高通量在线视觉检测研究。设计了一套群体鸡蛋图像高通量在线采集系统,其中运用Visual C++编写软件实现了上下位机的通讯及图像获取功能,使用STC89C52单片机接收光电开关的触发信号,共同配合实现了自动采集鸡蛋图像。在30000枚蛋/h六通道的传输装置上动态采集群体鸡蛋图像,采取较少但有效的预处理手段消除了高速传输对鸡蛋图像的影响,结合计算几何学中的凸包算法和最小二乘椭圆拟合重建鸡蛋外轮廓,解决了由于漏光引起蛋体图像凹陷现象的问题;通过分析长轴、短轴产生畸变的原因,对提取的长轴、短轴进行了修正处理,并建立长短轴像素点个数与实际测量尺寸的一元线性回归模型,其两者的相关系数分别为95.66%和94.39%,结合凸包算法相比于直接运用最小二乘椭圆拟合得到的相关系数更大,表明结合凸包算法的最小二乘椭圆拟合提取鸡蛋外形轮廓的精度更高。对84枚鸡蛋图像处理后进行验证,得到鸡蛋尺寸大小和外形扁平程度的分级准确率分别为90.5%和89.3%,单个鸡蛋的检测时间只需52.762ms,实现了鸡蛋尺寸形状的高通量在线检测分级。4)散黄蛋高通量在线视觉检测研究。为了进一步提高散黄蛋的检测效率,本研究动态采集15000枚蛋/h三通道传输装置上群体鸡蛋图像,首先利用与鸡蛋尺寸形状检测中相同的图像处理方法消除无用背景的干扰,获得仅含鸡蛋的目标图像;提取鸡蛋图像RGB空间和HSV空间的颜色分量平均值作为特征参数,分别利用随机森林RF和偏最小二乘判别PLS-DA建立散黄蛋分类模型,比较不同分类模型结果,发现利用RGB与HSV联合空间下的特征参数构建分类模型的效果最好,且RF分类模型优于PLS-DA。RGB与HSV联合空间下的散黄蛋RF分类模型预测识别率达到92.86%,单个鸡蛋的检测时间只需127.4ms,满足15000枚蛋/h高通量在线检测的要求,实现了高通量在线识别散黄蛋。5)破损蛋高通量在线视觉检测研究。在15000枚蛋/h三通道传输装置上动态采集群体鸡蛋图像,由于破损区域的位置具有随机性,因此单个鸡蛋需要通过综合采集三张图像的检测结果确定其是否破损。利用有效预处理方法获取鸡蛋目标图像,为了突显鸡蛋破损特征,使用了巴特沃斯高通滤波和灰度图像增强方法,但是同时也显现出斑点噪声区域;提取不同区域的形状特征参数(圆形度和最小外接矩形长宽比),建立粒子群PSO优化BP神经网络模型对破损区域和斑点噪声区域进行区分,区域类型识别率达到99.44%,表明PSO-BP-ANN模型相比于BP-ANN的泛化能力更好、鲁棒性更强。最后使用PSO-BP-ANN模型识别斑点噪声区域并予以消除,保留鸡蛋破损区域。对120枚鸡蛋进行验证,破损蛋识别率为91.67%,完好蛋识别率为95%,总体识别率达到93.33%,单枚鸡蛋的平均检测时间只需201.24ms,检测效率满足高通量在线检测的要求。
[Abstract]:Egg quality detection is the key link in egg commercialization, which is of great significance to improving the economic value of eggs and improving people's quality of life. In particular, high flux on-line detection has a great effect on improving the level of egg processing automation and the development of egg industry in China. Line rapid detection, this topic combined with the actual production requirements of egg processing, with egg freshness, scatter yellow, size shape, damage and other qualities as the focus of research, using spectral analysis and machine vision technology to detect the multiple qualities of eggs. The main research content and conclusion are as follows: 1) fast quantitative detection of egg freshness by optical fiber spectrometry Using the self built optical fiber spectral detection device to collect the transmission spectrum information of eggs, combined with Savitzky-Golay convolution smoothing filtering, multiple scattering correction, standard normal transformation, first order differential and two order differential pre processing methods, the partial least squares regression PLSR and support vector regression SVR model are established respectively, and the different model precision is compared. It is found that the SVR model of first order differential treatment has better prediction effect, and the SVR model is better than PLSR in general. It shows that SVR can extract the nonlinear relationship between egg freshness and spectral information. In order to simplify the quantitative model to detect the freshness of eggs quickly, the linear dimensionality reduction principal component analysis (PCA) and manifold learning are used. The nonlinear reduced dimension locally linear embedding LLE reprocessed the spectral data after the first order differential. The results of two dimensionality reduction were compared. The results showed that LLE better extracted the spectral effective information and improved the model accuracy. The reduction effect was more obvious than that of the PCA. The correlation coefficient and the root mean square error of the training set, the prediction set and the mean square error in the.LLE-SVR model were more obvious. Not for 0.922,7.21 and 0.911,8.80, cross validation the root mean square root error is lower than PCA-SVR, 0.79. research results show that LLE-SVR model can be used for rapid quantitative detection of egg freshness by optical fiber spectroscopy, a theoretical study of high throughput on-line detection of egg freshness in the future,.2) fast on-line identification of optical fiber spectra of scattered yellow eggs. The spectrum technique is used to dynamically collect the transmission spectrum data on 5000 egg /h single channel transmission devices. The wavelength optimization of different spectral preprocessed data is compared between the continuous projection algorithm SPA and the competitive adaptive weight weighting algorithm CARS respectively. It is found that the number of characteristic wavelengths selected by SPA is generally lower than that of CARS, and then the selected characteristic waves are combined. Using partial least squares discriminant PLS-DA, classification regression tree algorithm CART, K nearest neighbor classification algorithm KNN and cluster independent soft mode algorithm SIMCA four classifiers to establish multiple classifiers, according to the number of variables and recognition rate to compare the performance of the classifier, 5 classifiers are selected, and the detection time of single eggs is compared by each classifier. The SNV-SPA-PLS-DA model is suitable for online identification of yellow eggs, with only 3 characteristic variables, a single egg detection time of 55.733ms, a prediction accuracy of 97.14%, a high throughput on-line spectral identification of hellyellow eggs,.3) high throughput online visual detection of egg size and shape. A set of high pass group egg image high pass is designed. The on-line acquisition system is used, in which Visual C++ software is used to realize the communication and image acquisition function of the upper and lower computer. The trigger signal of the photoelectric switch is received by the STC89C52 single chip microcomputer, and the automatic collection of egg images is realized together, and the egg images are dynamically collected on the transmission and installation of 30000 egg /h six channels. But the effective preprocessing method eliminates the influence of high speed transmission on the egg image. Combined with the convex hull algorithm and the least square ellipse fitting in the calculation geometry to reconstruct the outer contour of the egg, the problem of the image sag caused by the leakage of the egg is solved. The reason of the distortion of the long axis and the short axis is analyzed, and the long axis and the short axis are extracted. The correction processing is carried out, and the linear regression model of the number of long and short axis pixels and the actual measurement size is established. The correlation coefficients of the two are 95.66% and 94.39% respectively. The correlation coefficient of the convex packet algorithm is larger than the least square ellipse fitting, which shows the least square ellipse fitting combined with the convex hull algorithm. The accuracy of the outline of egg shape is higher. After processing 84 eggs, the accuracy rate of the size and flat degree of egg is 90.5% and 89.3% respectively. The detection time of the single egg is only 52.762ms, and the high flux on-line detection and grading.4 of the egg size and shape is realized. In order to further improve the detection efficiency of the egg scattered, this study dynamically collected the image of the group egg on the 15000 /h three channel transmission device of egg. First, the interference of the useless background was eliminated by the same image processing method which was the same with the size shape detection of eggs. The target image containing only eggs was obtained, and the RGB space and HSV of the egg image were extracted. The average value of color component of space is used as the characteristic parameter. The classification model of scattered yellow eggs is established by using random forest RF and partial least squares discrimination PLS-DA respectively. The results of different classification models are compared. It is found that the best effect of using the characteristic parameters under the combined space of RGB and HSV is the best, and the RF classification model is superior to the joint space of PLS-DA.RGB and HSV. The prediction recognition rate of the RF classification model is 92.86%, the detection time of single egg is only 127.4ms, it meets the requirement of high throughput on-line detection of 15000 eggs /h, and the on-line visual detection of the high flux on the damaged egg with high throughput on-line identification of the yellow egg.5 is realized. The group chicken is dynamically collected on the 15000 egg /h three channel transmission device. Egg image, because the location of the damaged area is random, so the single egg needs to collect three images to determine whether the egg is damaged. Using the effective preprocessing method to obtain the egg target image, in order to highlight the egg breakage features, the Butterworth high pass filter and the gray image enhancement method are used, but the same method is used. It also shows the spot noise area, and extracts the shape feature parameters of different regions (circle degree and the minimum outer rectangle length width ratio), and establishes the particle swarm optimization BP neural network model to distinguish the damaged area and the spot noise region, and the region type recognition rate reaches 99.44%, indicating that the PSO-BP-ANN model is more generalization ability than the BP-ANN. In the end, the PSO-BP-ANN model is used to identify the spot noise area and to remove the damaged area of the egg. The identification rate of broken egg is 91.67%, the recognition rate of the egg is 95%, the overall recognition rate is 93.33%, the average detection time of the single egg is only 201.24ms, the detection efficiency meets the high flux. Requirements for online testing.
【学位授予单位】：华中农业大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TS253.7

【参考文献】