Image texture features as indicators of beef tenderness

Image processing techniques were developed to predict cooked-beef tenderness from fresh-beef image characteristics. Cattle from different finishing treatments were processed in a commercial plant. Two short loin steaks were sampled from each carcass; one used for sensory evaluation and the other for imaging. The samples varied significantly in both US quality grades and sensory tenderness scores. Color, marbling and texture features were extracted from the beef images. Statistical and neural network analyses were performed to relate the image features to sensory tenderness scores. Image texture features were found to be useful indicators of beef tenderness. Partial least squares and neural network models were able to predict beef tenderness from color, marbling and image texture features to R(2)-values up to 0.70.     

基于IGS-SVM模型的牛肉生理成熟度预测方法

生理成熟度是判定牛肉质量等级的重要指标,本实验建立一种通过改进的网格搜索（improved grid search,IGS）算法优化支持向量机（support vector machine,SVM）参数的模型,以实现牛肉的生理成熟度的预测。收集18、36、54、72 月龄的牛肉样本各25 个,共计100 个。利用机器视觉,采集样本的显微图像,经过图像处理后,提取不同生理成熟度牛肉的肌纤维特征参数,用统计学方法分析牛肉生理成熟度和肌纤维特征参数之间的相关性,并以肌纤维特征参数作为输入,利用76 个训练集样本,建立牛肉生理成熟度的SVM预测模型。为优化所建立的SVM模型,提出一种IGS算法,用其对SVM模型的约束参数C以及核函数参数g进行优化,结合留一交叉验证法得到最优的（C,g）参数组合,并将最佳参数代入分类器,得到优化的牛肉生理成熟度预测模型。用24 个测试集的独立样本检测模型的适用性并估测性能,结果表明：利用该模型对牛肉生理成熟度预测的准确率可达到91.67%;与传统网格搜索算法相比,IGS算法使得模型在训练时间上缩短了1 755.41 s。这表明所建立的模型具有较好的预测效果,也说明根据牛肉肌纤维的特征参数结合机器视觉及图像分析技术,对牛肉生理成熟度进行自动判定的方法是可行的。     

基于决策树雪花牛肉大理石花纹分级模型

为建立雪花牛肉大理石花纹等级评价方法,根据不同等级雪花牛肉大理石花纹图像特征及人工评级的标准,确定了影响大理石花纹的等级主要因素。本研究提出影响大理石花纹等级的几何参数特征、几何分布参数特征和统计参数特征。其中几何参数特征主要反映大理石花纹面积、周长等;几何分布特征主要反映大理石花纹图像中脂肪颗粒沉积的密度,根据脂肪颗粒沉积情况可分为大颗粒脂肪、中颗粒脂肪、小颗粒脂肪等;统计参数特征主要反映大理石花纹丰富程度以及大理石花纹分布均匀性。利用相关性分析提取影响雪花牛肉大理石花纹等级的特征参数。建立基于C4.5和CART算法的决策树模型,结果表明：对于C4.5算法建立的决策树分级模型,三级和五级大理石花纹分级预测精度分别为91.80%、92.31%,而该模型针对四级样本建立的模型无效,其结果多数误判为三级;对于CART算法建立的决策树模型同样存在这样的问题,即三级和五级大理石花纹分级预测精度高,而对四级样本分级无效。     

基于高光谱和超声成像技术的原切与合成调理牛排鉴别

针对市场上存在合成调理牛排冒充原切售卖的现象,研究利用高光谱和超声成像技术对它们进行鉴别的方法。分别采集原切与合成调理牛排的高光谱和超声图像信息,利用灰度共生矩阵法提取图像的纹理特征值,分别建立线性判别分析、K最邻近（K-nearest neighbor,KNN）、反向传播人工神经网络和极限学习机（extreme learning machine,ELM）4?种鉴别模型,而后将2?种技术数据融合建模,并采用连续投影法、竞争性自适应重加权算法（competitive adaptive reweighted sampling,CARS）、变量组合集群分析（variables combination population analysis,VCPA）法3?种方法筛选特征变量建模。结果表明：合成调理牛排的肉块组织均匀,超声图像信号弱、均一性好,与原切调理牛排图像存在差异。高光谱和超声成像技术的最佳模型分别为KNN和ELM,模型预测集识别率分别为95.00%和90.00%。数据融合后建模,最佳模型ELM模型预测集识别率模型为97.50%,在3?种变量选择方法中,CARS和VCPA选择的纹理变量建立的模型预测集识别率达到100.00%。研究表明高光谱和超声成像数据融合结合变量选择方法可以快速准确地鉴别原切和合成调理牛排。     

Evaluation of pork color by using computer vision

The objective of this study was to determine the potential of computer vision technology for evaluating fresh pork loin color. Software was developed to segment pork loin images into background, muscle and fat. Color image features were then extracted from segmented images. Features used in this study included mean and standard deviation of red, green, and blue bands of the segmented muscle area. Sensory scores were obtained for the color characteristics of the lean meat from a trained panel using a 5-point color scale. The scores were based on visual perception and ranged from 1 to 5. Both statistical and neural network models were employed to predict the color scores by using the image features as inputs. The statistical model used partial least squares technique to derive latent variables. The latent variables were subsequently used in a multiple linear regression. The neural network used a back-propagation learning algorithm. Correlation coefficients between predicted and original sensory scores were 0.75 and 0.52 for neural network and statistical models, respectively. Prediction error was the difference between average sensory score and the predicted color score. An error of 0.6 or lower was considered negligible from a practical viewpoint. For 93.2% of the 44 pork loin samples, prediction error was lower than 0.6 in neural network modeling. In addition, 84.1% of the samples gave an error lower than 0.6 in the statistical predictions. Results of this study showed that an image processing system in conjunction with a neural network is an effective tool for evaluating fresh pork color.     

Analysis of image-based measurements and USDA characteristics as predictors of beef lean yield

This paper describes a comprehensive analysis of the 12th rib image measurements, USDA yield characteristics and USDA yield grade as predictors of beef lean yield. The predictors were used in separate groups to construct three sets of multiple linear regression models for the prediction of lean yield of 241 carcasses. Fat thickness is traditionally considered the most useful predictor. The analysis showed that the percent rib eye area was a more useful single predictor than fat thickness, and that the average fat thickness was insignificant when rib eye area and fat area were used. While marbling is traditionally considered a predictor of quality, the results showed that marbling characteristics were also useful for yield prediction. The usually-observed large differences in accuracy between the predictions of lean weight and lean percentage were shown to result from the variations in carcass weight. Statistical diagnostics confirmed the suitability of the models developed. A neural network model was tested and the results suggested that the inclusion of nonlinearity in the predictive models did not prove beneficial. Important predictors were identified and the advantages of computer vision and image processing techniques were further demonstrated.     

基于BPNN和SVM的烟叶成熟度鉴别模型

为了在烟叶采收环节中快速准确地鉴别适熟烟叶。以下部烟叶为样本,利用烟叶图像的RGB颜色特征信息,以目标区域和背景区域平均灰度值的最大对比度为依据,选取G-B图像提取烟叶目标区域。目标区域通过形态学膨胀和腐蚀运算去除噪声后,从原始烟叶图像中得到目标烟叶图像。通过MATLAB软件提取R、G、H、S、V颜色特征均值,以及能量ASM、熵ENT、惯性矩INE、相关性CORRL纹理特征,将这9个变量作为输入参数,分别建立基于BP神经网络、支持向量机的烟叶成熟度鉴别模型,准确率分别为93.83%和97.53%。试验结果表明,通过机器视觉对烟叶成熟度鉴别是可行的,为进一步研制烟叶采收机奠定了基础。      

USDA MATURITY INDICES AND PALATABILITY OF BEEF RIB STEAKS1

In a population of 195 beef carcasses, as maturity increased (by whole USDA groups) from A to B, flavor, tenderness and overall palatability ratings for rib steaks decreased (P < .05); further increases in maturity (B vs. E) resulted in subsequent decreases (P < .05) in tenderness and overall palatability. Subdivision of whole maturity groups into thirds and subsequent data analyses supported precepts in USDA grade standards of declining palatability with advancing maturity, but the declines, when they occurred, were nearly linear through the extremes (from A^? to E⁺) and did not support the concept, in USDA grade standards, of a precipitous decline in eating satisfaction at or near the B⁺/C^? maturity line. Among seven carcass maturity indices, all were similarly and singularly related (P < .05) to palatability of rib steaks, with some measure of skeletal maturity (the best of these was color/shape of the rib bones) plus some measure of muscle maturity (the best of these was color of the longissimus muscle) combining (by means of multiple regression analyses) to predict tenderness (panel ratings and shear values) with maximum precision. Neither total pigment concentration (a chemical measure of muscle color), amount of ether-extractable lipid (a chemical measure of intramuscular fatness) nor marbling score (a visual measure of intramuscular fatness) was able to explain more than about 7% of the observed variability in palatability ratings for cooked beef steaks. Data suggest that more attention should be paid to appearance of rib bones in a carcass (to better assess effects of maturity) and that less attention could be paid to differences between SLIGHT, SMALL and MODEST in marbling in the ribeyes, in attempting to predict, via USDA grading, the palatability of beef rib steaks.     

Utilization of image processing to quantitate surface metmyoglobin on fresh beef

Ground beef patties were manufactured with various combinations of ascorbic acid and mechanically recovered neck bone lean (MRNL) to study the use of image processing in predicting percentage surface metmyoglobin (metMb) on fresh beef. Ascorbic acid and MRNL cause various color phenomena that resulted in a wide range of variation in surface color. Patties were also stored over six days of retail display to cause further color changes. Surface color was assessed by several different accepted methods. A prediction equation for percentage surface metMb included mean values for hue, saturation and intensity. Root mean square error, R-square and Mallow's Cp statistic were used as selection criteria for choosing the best predictive model. Image processing hue, saturation and intensity accounted for 93% of the variation in percentage surface metMb. Since hue, saturation and intensity each contribute to overall color profile it is logical that these parameters are good predictors. These data indicate that image processing is capable of objectively measuring percentage surface metMb.     

采用多蜂群协同演化算法的服装流行色预测

针对现有服装流行色定量预测方法存在的精度缺陷,借鉴了复杂适应系统理论和生物学协同进化思想,提出一种基于层次协同演化机制的多蜂群协同优化算法。将该算法应用于人工神经网络权值配置问题上,通过拟合曲线进行测试。结果表明该算法可提升神经网络模拟目标问题的精度。使用 PANTONE 发布的2007−2016 年的国际春夏女装流行色定案,建立了春夏女装流行色色相的预测模型;对该模型中改进的神经网络进行训练,分析不同隐含层节点个数对色相预测精度的影响;预测2017 年女装流行色色相,并将预测结果与PANTONE 官方发布结果进行对比。结果表明该方法与其他方法相比提高了预测结果的精准度。     

基于机器视觉的马铃薯加工原料分选系统

目的：设计以马铃薯加工原料为对象的自动化分选系统,实现设定标准下马铃薯自动识别。方法：构建分选系统的控制流程及分选算法,通过自动传送、机器视觉采集、吸压翻转自动化获取马铃薯2面的图像,采用图像复原算法消除运动模糊,设计面积比、长短径、凸起检测算法对马铃薯畸形、发芽、大小进行检测,基于颜色特征构建神经网络模型对马铃薯绿皮、病斑、常色进行分类。结果：利用BP神经网络算法预测马铃薯外观颜色绿皮、病斑、正常的分类,以误差分数为衡量预测模型准确性的度量,神经网络的预测分类平均准确率为96.2%。通过选取混合样本对分选系统进行测试,参照设定分选标准,分选系统对马铃薯识别正确率达到95.92%;单薯处理耗时3.76 s。系统运行稳定。结论：该方法用于马铃薯加工原料精量分选可行,能够满足薯制品加工生产线前端的分选需要。     

基于高光谱图像光谱变量和颜色特征的霉变玉米籽粒识别

目的:准确识别霉变玉米籽粒。方法:基于高光谱图像光谱变量和颜色特征建立霉变玉米籽粒识别的新方法。先对玉米籽粒图像进行图像分割和光谱变量、颜色特征提取,并根据颜色特征生成颜色直方图;将光谱变量和颜色直方图特征组成特征集合;通过距离函数对特征集合中所有特征的分析确定霉变玉米籽粒所属类别。结果:所提方法对霉变玉米籽粒类别的最大平均识别偏差为1.12,最佳平均识别准确率为97.59%;与基于高光谱图像+随机蛙跳+极限学习机的方法、基于高光谱图像+稀疏自动编码器+卷积神经网络的方法、基于高光谱图像+蚁群优化+BP神经网络的方法相比,研究所提方法对霉变玉米籽粒类别的识别准确率明显提高。结论:该方法可实现被测玉米籽粒样品是否霉变以及霉变程度的准确判断。     

基于颜色和纹理特征的图像情感语义分类

利用图像的低层特征实现了图像高层情感语义(happy和sad)的分类:通过在HSV颜色空间中提取图像的全局颜色特征,并利用黄金分割原理提取位于视觉中心位置主要区域的局部颜色特征,结合二维Gabor小波变换提取全局图像的纹理特征,实现对自然风景图像进行情感特征提取.采用PCA方法对情感特征进行降维,将降维后的特征向量结合BP神经网络,完成情感语义分类检索.     

DETERMINING FAT CONTENT IN GROUND BEEF VIA COLOR IMAGE PROCESSING

Two lots of beef trim were formulated into nine meat blocks which ranged from 9 to 32% fat. Ground samples (n=2) were randomly taken from each meat block, images were captured in the hue, saturation and intensity (HSI) format and fat content were determined. Patties from each block were manufactured and randomly selected (n=3) for image capture and fat content determination. The HSI statistical function characteristics for the color distribution curves were evaluated. Fat content was correlated to many of the image characteristics (p < 0.05). Image variables accounted from 81 to 87% of the variation in actual fat content, depending on which form was analyzed, and residual standard deviations ranged from 2.75 to 3.62%. These data show that image processing may be a useful technology for monitoring ground beef composition with fat levels up to 32%.     

基于混合色彩空间独立特征的色纺面料颜色表征模型

为解决色纺加工中染色纤维整体呈色特性难以准确描述的问题,建立了基于混合色彩空间的独立特征颜色分析模型。对Lab与HSV 2种颜色空间中具有相同属性的颜色分量进行独立融合,并构建混合色彩空间;在此基础上,分别采用三阶颜色矩和局部纹理统计特征对色纺织物图像的颜色信息进行表征与融合。结果表明:对于色纺针织物或梭织物而言,所建立的颜色表征模型不仅能够对较大范围内的质量配比变化进行有效表征,而且对于染色纤维细微调整而导致的颜色改变亦能准确表达,具有理想的鲁棒性与普适性。     

基于LCH视觉频率响应特性的图像编码研究

对比敏感度函数(CSF)是表征人眼视觉频率响应的重要方法之一。本研究以LCH颜色空间为基础,结合大量实验构建出的CSF模型曲线,生成基于多种中心色调角的明度、彩度和色调角滤波器,根据阈值限定,对图像进行对应的滤波压缩处理,并编制出用于图像处理的C语言程序。根据处理后得到的图像,将基于LCH-CSF模型的图像与主流JPEG压缩图像进行对比评价,为图像压缩方法的研究提供了新的思路,同时也验证了LCH颜色空间下所构建的CSF模型曲线与人眼视觉频率响应的契合程度。     

Image Processing Applied to Classification of Avocado Variety Hass (Persea americana Mill.) During the Ripening Process

This work was undertaken to analyze the ripening process of avocados variety Hass (Persea americana Mill.) by image processing (IP) methodology. A set of avocados (10 samples) was used to follow the changes in image features during ripening by applying a computer vision system, extracting color and textural parameters. Other 16 avocados were used to evaluate the firmness and mass loss. Three maturity stages of avocados were established, and a classification was obtained by applying principal component analysis and k-nearest neighbor algorithm. During the ripening process (12 days), avocado firmness decreased from 75.43 to 2.63 N, while skin color values kept invariable during 6 days; after that, a decrement in the peel green color (a*) was observed (−9.68 to 2.32). Image features showed that during ripening the color parameters (L*, a*, and b*), entropy (4.29 to 4.00), angular second moment (0.287 to 0.360), and fractal dimension (2.58 to 2.44) had a similar path as compared to mass loss, a*, and firmness ripening parameters, respectively. Relationships between image features and ripening parameters were obtained. The parameter a* was the most useful digital feature to establish an acceptable percentage of avocado classification (>80%) in three different maturity stages found. Results obtained by means of IP could be useful to evaluate, at laboratory level, the ripening process of the avocados.     

基于不变矩和改进SVM的牛肉大理石纹评级

针对牛肉大理石纹人工评级过程中人为误差干扰大的问题,研究利用图像处理技术提高牛肉大理石纹评级的客观性并增强自动化程度,提出基于不变矩、灰度共生矩阵和混沌蜂群优化混合核函数支持向量机(Support Vector Machine,SVM)的牛肉大理石纹评级法。首先计算牛肉大理石纹图像的不变矩和灰度共生矩阵统计量并由此构造特征向量;然后将训练和测试样本输入到混合核函数SVM,采用混沌蜂群算法优化SVM中的惩罚因子及核参数,使其分类识别性能达到最优;最后输入待评级样本进行分类识别,得到最优评级结果。大量实验结果表明:根据我国NY/T676-2010牛肉大理石纹标准图谱,评级正确率分别达到100%(一级)、93.3%(二级)、93.3%(三级)、96.7%(四级)、100%(五级)。与基于灰度矩和SVM法、基于灰度共生矩阵和BP(Back Propagation)神经网络法相比,本文所得评级正确率最高,且与专业评级师的实际评级情况最相符。     

Consumer Assessment of Beef Tenderloin Steaks from Various USDA Quality Grades at 3 Degrees of Doneness

A consumer study was conducted to determine palatability ratings of beef tenderloin steaks from USDA Choice, USDA Select, and USDA Select with marbling scores from Slight 50 to 100 (USDA High Select) cooked to various degrees of doneness. Steaks were randomly assigned to 1 of 3 degree of doneness categories: very‐rare, medium‐rare, or well‐done. Consumers (N = 315) were screened for preference of degree of doneness and fed 4 samples of their preferred doneness (a warm‐up and one from each USDA quality grade treatment in a random order). Consumers evaluated steaks on an 8‐point verbally anchored hedonic scale for tenderness, juiciness, flavor, and overall like as well as rated steaks as acceptable or unacceptable for all palatability traits. Quality grade had no effect (P > 0.05) on consumer ratings for tenderness, juiciness, flavor, and overall like scores, with all traits averaging above a 7 (“like very much”) on the 8‐point scale. In addition, no differences (P > 0.05) were found in the percentage of samples rated as acceptable for all palatability traits, with more than 94% of samples rated acceptable for each trait in all quality grades evaluated. Steaks cooked to well‐done had lower (P < 0.05) juiciness scores than steaks cooked to very‐rare or medium‐rare and were rated lower for tenderness (P < 0.05) than steaks cooked to a very‐rare degree of doneness. Results indicate consumers were not able to detect differences in tenderness, juiciness, flavor, or overall like among beef tenderloin steaks from USDA Choice and Select quality grades.     

基于多源信息融合的绿茶杀青叶水分含量智能感知方法

为了实现绿茶杀青过程中水分含量的快速有效检测,利用机器视觉结合近红外光谱技术,构建绿茶杀青过程中水分含量变化的定量预测模型。首先采集杀青过程中在制品的光谱和图像信息,然后采用竞争性自适应权重取样（competitive adaptive reweighted sampling,CARS）法、变量组合集群分析（variables combination population analysis,VCPA）法、变量组合集群分析法结合迭代保留信息变量（variable combination population analysis and iteratively retains informative variables,VCPA-IRIV）法和随机蛙跳法（random frog,RF）4 种变量筛选方法提取光谱中的特征波长,并融合图像中的15 个色泽和纹理特征建立线性偏最小二乘回归（partial least squares regression,PLSR）和非线性支持向量回归（support vector regression,SVR）预测模型。结果表明,与单一数据相比,基于融合数据所建立的模型能有效提高预测精度,其中基于CARS算法提取光谱特征波长融合图像的15 个颜色特征,并结合归一化预处理和主成分分析（principal component analysis,PCA）建立的SVR模型效果最佳,其中校正集相关系数为0.974 2,预测集相关系数为0.971 9,相对分析误差（relative percent deviation,RPD）为4.154 6,表明模型具有极好的预测性能。综上,本研究证明融合光谱和图像技术对绿茶杀青过程中水分含量预测的可行性,克服了单一传感器预测精度低的问题,为实现绿茶杀青叶水分含量的快速无损检测和精准把控杀青质量提供理论基础。