【作者】 于洋；

【导师】 田海清；

【作者基本信息】 内蒙古农业大学 ， 农业机械化工程， 2021， 硕士

【摘要】 嫩度是羊肉品质评价的重要指标之一,它影响着肉品的口感以及商业价值。传统羊肉嫩度检测方法效率低、破坏样品,难以满足目前肉品品质快速无损检测的需求,而基于单一光谱或图像特征信息建立的模型无法全面准确地对肉品品质进行评价,所以,寻找更为精确和高效的方法显得尤为重要。近些年,高光谱成像技术在食品检验方面得到广泛应用,同时也是研究热点。另外偏振图像中包含丰富的纹理信息,故偏振成像技术为肉类品质快速无损检测提供了新方法。因此,为探究冷鲜羊肉嫩度快速无损检测方法,本文以内蒙古锡林郭勒羊肉为研究对象,基于特征层信息融合方法建立了冷鲜羊肉嫩度预测模型,实现了对冷鲜羊肉嫩度的快速无损检测。具体研究内容与成果如下:(1)以120块冷鲜羊肉为研究对象,利用高光谱成像系统获取不同贮藏天数的羊肉高光谱图像数据,依据NY/T1180-2006标准测定羊肉嫩度,提取羊肉图像反射光谱曲线,采用多元散射校正法对原始光谱反射率进行校正。在此基础上,利用主成分分析法确定了羊肉嫩度的特征波长,分别为620.23nm、761.48nm、819.48 nm,并提取了对应波长下的灰度图像。(2)对特征图像进行灰度共生矩阵(CLCM)纹理提取,分别基于特征波长信息、特征图像纹理信息以及图谱特征层融合信息建立冷鲜羊肉嫩度的BP神经网络和支持向量机(SVM)预测模型。结果表明,BP神经网络模型精度总体高于SVM预测模型,基于图谱特征层融合信息建立的BP和SVM模型预测集决定系数R~2分别为0.8527和0.7964,均方根误差RMSEP分别为1.7623和2.1541。(3)利用偏振成像系统采集冷鲜羊肉样本图像,利用随机筛选的90个样本建立预测模型。采用LBP算法进行纹理特征提取,经8次试验后确定最优采样像素点P和半径参数R分别为16和2,建立的BP神经网络和SVM模型决定系数R~2分别为0.8212和0.7641,均方根误差RMSEC分别为2.6581和2.7821。采用CLCM提取偏振图像纹理特征并建立BP神经网络模型和SVM模型,其决定系数R~2分别为0.8099和0.7708,RMSEC分别为2.7296和2.7569。将CLCM和LBP信息特征层进行融合形成新纹理特征,并建立BP和SVM模型,其决定系数R~2分别为0.8318和0.7932,RMSEC分别为2.5837和2.8223。对比模型预测精度发现,基于特征层融合所建立的BP、SVM预测模型验证集决定系数R~2分别为0.8100、0.7685,RMSEP分别为2.0712、1.9093,表明CLCM和LBP特征融合后的模型精度高于单一信息模型精度。综上所述,上述结果验证了高光谱成像技术和偏振成像技术在羊肉嫩度品质检测方面的有效性,为羊肉快速无损检测提供了新思路。更多还原

【Abstract】 Tenderness is one of the important indicators of lamb quality evaluation,it affects the taste and commercial value of meat.Traditional lamb tenderness detection methods have low efficiency and damage samples,which are difficult to meet the current needs for fast and non-destructive testing of meat quality.However,models based on single spectrum or image feature information cannot comprehensively and accurately evaluate meat quality.Therefore,look for more It is particularly important for precise and efficient methods.In recent years,hyperspectral imaging technology has been widely used in food inspection,and it is also a research hotspot.In addition,the polarization image contains rich texture information,so the polarization imaging technology provides a new method for fast and non-destructive inspection of meat quality.Therefore,in order to explore the rapid non-destructive testing method of cold fresh lamb tenderness,this thesis takes Xilin Gol lamb in Inner Mongolia as the research object,and establishes a cold fresh lamb tenderness prediction model based on the feature layer information fusion method,which realizes the rapid and non-destructive detection of cold fresh lamb tenderness.Detection.The specific research content and results are as follows:(1)Taking 120 pieces of cold fresh lamb as the research object,the hyperspectral imaging system was used to obtain hyperspectral image data of lamb stored for different days,the tenderness of lamb was measured according to the NY/T1180-2006 standard,and the reflectance spectrum curve of lamb image was extracted.The scatter correction method corrects the original spectral reflectance.On this basis,the characteristic wavelengths of lamb tenderness were determined by principal component analysis,which were 620.23nm,761.48nm,and 819.48nm,and the gray-scale images at the corresponding wavelengths were extracted.(2)Extract the gray-level co-occurrence matrix texture of the feature image,and build the BP neural network and support vector machine(SVM)prediction model of cold fresh lamb tenderness based on the feature wavelength information,feature image texture information and map feature layer fusion information.The results show that the accuracy of the BP neural network model is generally higher than that of the SVM prediction model.The determination coefficients R~2of the BP and SVM model prediction sets based on the fusion information of the map feature layer are 0.8527 and 0.7964,respectively,and the root mean square error RMSEP is 1.7623 and 2.1541,respectively.(3)Use the polarization imaging system to collect images of cold fresh lamb samples,and use 90 randomly selected samples to establish a prediction model.LBP algorithm is used for texture feature extraction.After 8 experiments,the optimal sampling pixel point P and radius parameter R are determined to be 16 and 2,respectively.The established BP neural network and SVM model determination coefficients R~2are 0.8212 and 0.7641,respectively,root mean square The error RMSEC are 2.6581 and 2.7821 respectively.The gray-level co-occurrence matrix is used to extract the texture features of the polarization image,and the BP neural network model and the SVM model are established.The determination coefficients R~2are 0.8099 and 0.7708,and the RMSEC are 2.7296 and2.7569,respectively.The CLCM and LBP information feature layers are fused to form new texture features,and the BP and SVM models are established.The determination coefficients R~2are 0.8318 and 0.7932,respectively,and the RMSEC are 2.5837 and 2.8223,respectively.Comparing the model prediction accuracy,it is found that the determination coefficients R~2of the verification set of the BP and SVM prediction models established based on the feature layer fusion are 0.8100 and 0.7685,respectively,and the RMSEP are2.0712 and 1.9093,respectively,indicating that the model accuracy after the fusion of CLCM and LBP features is higher than that of single information Model accuracy.In summary,the above results verify the effectiveness of hyperspectral imaging technology and polarization imaging technology in the quality detection of lamb tenderness,and provide a new idea for rapid non-destructive testing of lamb.更多还原