CLASSIFICATION OF ON-LINE POULTRY CARCASSES WITH BACKPROPAGATION NEURAL NETWORKS1

A transportable system equipped with an overhead shackle conveying line and a visible/near-infrared (Vis/NIR) spectrophotometer system was assembled and used at a poultry slaughter plant. The reflectance spectra of each poultry carcass hung on the moving shackle was measured with a stationary fiber optic probe, which was set 2 to 5 cm away from the carcass, depending on the size. Reflectance spectra of wholesome and unwholesome poultry carcasses on the moving shackle, set at 60 or 90 birds/min, were measured, either under room light or in a dark environment. The scanning time for each carcass was 0.32 s. Most of the unwholesome poultry carcasses for this study were septicemic and air-sacculitic. The average accuracy in classifying wholesome and unwholesome carcasses was above 94%. All the misclassified carcasses were air-sacculitic. With a shackle speed of 90 birds/min, the highest average accuracy was obtained when the reflectance was measured in the dark (97.5%). The results showed that the accuracy of classification could be improved with the maintenance of a consistent lighting environment. All results indicated the Vis/NIR spectrophotometer system would be a highly accurate, robust tool for on-line, real-time classification of wholesome and unwholesome carcasses.     

A CHICKEN CARCASS INSPECTION SYSTEM USING VISIBLE/NEAR‐INFRARED REFLECTANCE: IN‐PLANT TRIALS1

On‐line trials of an industrial prototype visible/near‐infrared spectrophotometer system developed by the Instrumentation and Sensing Laboratory for inspecting poultry for diseased and defective carcasses were conducted during an 8‐day period in a slaughter plant in New Holland, Pennsylvania. Spectra (470–960 nm) of 1174 normal and 576 abnormal (diseased and/or defective) chicken carcasses were measured. The instrument measured the spectra of veterinarian‐selected carcasses as they passed on a processing line at a speed of 70 birds per minute. Classification models using principal component analysis as a data pretreatment for input into neural networks were able to classify the carcasses from the spectral data with a success rate of 95%. Data from 3 days can predict the subsequent two days’chickens with high accuracy. This accuracy was consistent with the results obtained previously in off‐line studies. Thus, the method shows promise for separation of diseased and defective carcasses from wholesome carcasses in a partially automated inspection system. Details of the models using various training regimens are discussed.     

REAL‐TIME DUAL‐WAVELENGTH IMAGE PROCESSING FOR POULTRY SAFETY INSPECTION1

Real‐time multispectral image processing algorithms were developed for online poultry carcass inspection. Neural network models with different learning rules (delta and hyperbolic tangent) and transfer functions (sigmoid and norm‐cum‐sigmoid) were examined using features extracted from spectral images at 540 nm and 700 nm. The classification accuracy using dual wavelength spectral images was much higher than single wavelength spectral images in identifying unwholesome poultry carcasses. The spectral image features at 700 nm were useful to identify wholesome carcasses, while the combination of spectral image features at 540 nm, 700 nm, and their ratio improved the classification accuracy of unwholesome carcasses. The optimum neural classifier utilized delta learning rule and hyperbolic tangent transfer function. The classification accuracy was 91.1% for wholesome and 83.3 % for unwholesome carcasses when the spectral images of all 540 nm, 700 nm, and their ratio were used as inputs to the neural network model.     

Combination and comparison of multivariate analysis for the identification of orange varieties using visible and near infrared reflectance spectroscopy

A study of multivariate analysis for orange varieties was carried out, and the potential of visible and near infrared reflectance spectroscopy (Vis/NIRS) for its ability to nondestructively differentiate orange varieties was evaluated. A total of 320 orange samples (80 for each variety) were investigated for Vis/NIRS on 325–1075 nm using a field spectroradiometer. Multivariate classification methods including principal component analysis (PCA), back propagation neural network (BPNN) and partial least squares discriminant analysis (PLSDA) were adopted to classify oranges. Sixteen principal components from PCA were used as the input of BPNN model, and the identification accuracy of four orange varieties reached 100%. The prediction result of PLSDA, i.e., standard error of prediction (SEP) 0.24497, correlation coefficient (R) 0.97843, root mean square error of prediction (RMSEP) 0.24268, and identification accuracy 90% indicate that PLSDA is an alternative model for orange identification. With the comparison of these two models, it shows that BPNN combined with PCA obtained better classification effect than that of PLSDA. The overall results demonstrate that Vis/NIRS technology with multivariate analysis models is promising for the rapid and reliable determination for identification of orange varieties.     

Storage Stability of Chicken as Affected by Map and Lactic Acid Treatment

The effects of MAP and/or pretreatment with lactic acid buffer were studied on the shelf-life of broiler carcasses, under conditions simulating market storage (4 and 7°C) in the State of Kuwait. Pretreatment with lactic acid buffer increased treated carcass shelf-life by 6–7 days when stored at 4°C and 5–6 days at 7°C. Likewise, pretreatment of MAP carcasses with the same buffer extended the shelf-life to >36 and 35 days, compared with 22 and 13 days for the untreated MAP carcasses. Pretreatment of poultry with lactic acid buffer, with or without MAP, provides a potential alternative for improving the storage quality of poultry.     

基于卷积神经网络的领带花型情感分类

为避免传统手工特征和局部特征难以全面表征和准确量化图像情感特征的不足,以领带花型为研究对象,提出了一种融合手工情感特征的基于卷积神经网络的织物图像情感分类方法,可为服饰设计、服装选购等提供辅助。 首先对领带花型图像进行情感评价,建立领带花型图像的情感样本库;然后提取图像饱和度、纹理等手工情感特征和图像像素值作为卷积神经网络的输入;其次建立卷积神经网络模型,将2 000幅样本图像作为训练样本对卷积神经网络进行训练;最后将1 000幅检测样本输入训练后的卷积神经网络,实现了领带花型图像的情感分类。实验结果显示：该方法的情感分类准确率为89.7%,比采用传统手工特征的分类方法有较大提升,较其他卷积神经网络模型正确率更高。     

Fast line-scan imaging system for broiler carcass inspection

The USDA Agricultural Research Service has developed a fast line-scan imaging system for differentiating wholesome and systemically diseased fresh chickens. The imaging system was used to acquire hyperspectral line-scan images of 250 chicken carcasses on a laboratory processing line moving at 70 birds per minute. A method appropriate for line-scan imaging was developed for automated sensing of birds and locating the Region of Interest (ROI) within the line-scan images most suited for differentiation. From analysis of wholesome and systemically diseased chicken spectra in the ROI, four key wavelengths for differentiating between wholesome and systemically diseased chickens were selected: 424, 465, 515, and 546 nm. The key wavelengths and their ratios with a reference wavelength (689 nm) were investigated for a fuzzy logic based differentiation algorithm. Classification using the key wavelengths correctly identified 98 and 95% of wholesome and systemically diseased chickens for model development, and 98 and 93% of wholesome and systemically diseased chickens for model testing. Although band ratios reduced variation within each chicken category, the resulting classification accuracies were not significantly improved over those for classification by key wavelengths.     

基于非靶向风味组学分析3 种品牌啤酒的风味差异

通过顶空固相微萃取结合气相色谱-质谱鉴定Lager啤酒中挥发物,并利用主成分分析、偏最小二乘-判别分析（partial least squares-discrimination analysis,PLS-DA）法及随机森林分类（random forest classifier,RFC）法区分3 种品牌的Lager啤酒的风味差异。使用二乙烯基苯/碳分子筛/聚二甲基硅氧烷萃取纤维,室温固相微萃取5 min,然后在气相色谱仪进样口解吸,并通过色谱分离进行啤酒挥发物分析。在被鉴定的风味化合物中,辛酸乙酯、癸酸乙酯等是区分Lager啤酒品牌的关键性风味物质。在建立的PLS-DA与RFC模型中上述组分对啤酒风味差异的贡献较大,模型的预测准确度均达到100%,此策略对基于风味的啤酒分类方法的建立起到积极作用。     

MULTISPECTRAL IMAGING SYSTEM FOR FECAL AND INGESTA DETECTION ON POULTRY CARCASSES

A multispectral imaging system, including a common aperture camera with three optical trim filters (515.4, 566.4 and 631 nm), which were selected by visible/near‐infrared NIR spectroscopy and validated by a hyperspectral imaging system, was developed for a real‐time, on‐line poultry inspection application. The algorithm developed by a hyperspectral imaging system was employed for multispectral image analysis to validate the accuracy of fecal and ingesta detection in real‐time poultry processing. A multispectral imaging system could be used effectively for detecting feces (from duodenum, ceca, and colon) and ingesta on the surface of poultry carcasses with the processing time of approximately 251 ms or 3.99 frames/s. The multispectral imaging system developed in this research can be used for real‐time, on‐line detection of fecal and ingesta contaminant on poultry carcasses. The overall accuracy to identify fecal and ingesta contaminants was 96.8% and the prediction accuracy to identify each contaminant were 92.4% for duodenum and 98% for ceca, colon, and ingesta with moderate false positives.     

Detection of Fecal Residue on Poultry Carcasses by Laser-Induced Fluorescence Imaging

ABSTRACT:  Feasibility of fluorescence imaging technique for the detection of diluted fecal matters from various parts of the digestive tract, including colon, ceca, small intestine, and duodenum, on poultry carcasses was investigated. One of the challenges for using fluorescence imaging for inspection of agricultural material is the low fluorescence yield in that fluorescence can be masked by ambient light. A laser-induced fluorescence imaging system (LIFIS) developed by our group allowed acquisition of fluorescence from feces-contaminated poultry carcasses in ambient light. Fluorescence emission images at 630 nm were captured with 415-nm laser excitation. Image processing algorithms including threshold and image erosion were used to identify fecal spots diluted up to 1: 10 by weight with double distilled water. Feces spots on the carcasses, without dilution and up to 1: 5 dilutions, could be detected with 100% accuracy regardless of feces type. Detection accuracy for fecal matters diluted up to 1: 10 was 96.6%. The results demonstrated good potential of the LIFIS for detection of diluted poultry fecal matter, which can harbor pathogens, on poultry carcasses.     

Prediction of food thermal process evaluation parameters using neural networks

Two neural networks (ANN) were developed to predict thermal process evaluation parameters g and f(h)/U (the ratio of heating rate index to the sterilizing value), respectively. The temperature change required for the thermal destruction curve to traverse one log cycle (z), cooling lag factor (j(c)) andf(h)/U were input variables for predicting g and z, while j(c) and g were inputs for predicting f(h)/U. The data used to train and verify the ANN were obtained from reported values. Shrinking of input and output variables using natural logarithm function improved the prediction accuracy. The use of "Wardnets" with three slabs of 14 nodes in each slab, with a learning rate of 0.7 and momentum of 0.9 provided the best predictions. The g (unshrunk values) was predicted with a mean relative error of 1.25 +/- 1.77%, and a mean absolute error of 0.11 +/- 0.16 degrees F. The f(h)/U was predicted with a mean relative error of 1.41 +/- 3.40%, and a mean absolute error of 2.43 +/- 15.97, using 10 nodes in each slab. The process time calculated using the g from the ANN models closely followed the time calculated from the tabulated gvalues (RMS=0.612 min, average absolute error=0.466 min with an S.D. of 0.400 min).     

基于嗅觉可视化技术的食用植物油分类识别

为实现山茶油与3种常见食用植物油(菜籽油、大豆油和玉米油)的区分,制备可视化传感器阵列,采用嗅觉可视化技术对4种不同种类的食用植物油进行分类识别。采用主成分分析(PCA)对4种油样的特征数据进行降维,然后将降维后的数据导入K近邻(KNN)、极限学习机(ELM)、支持向量机(SVM) 3种分类模型中进行模型参数优化,对比了3种分类模型的分类结果。结果表明：建立的SVM分类模型性能最优,当输入主成分向量数为7、c=1.741 1、g=4.549 8时,SVM分类模型的测试集分类识别准确率为95.8%,五折交叉验证准确率为89.6%。制得的可视化传感器阵列可以实现4种食用植物油的分类识别,嗅觉可视化技术用于食用植物油检测是可行的。     

Real-time multispectral imaging system for online poultry fecal inspection using unified modeling language

A prototype real-time multispectral imaging system for fecal and ingesta contaminant detection on broiler carcasses has been developed. The prototype system includes a common aperture camera with three optical trim filters (517, 565 and 802-nm wavelength), which were selected by visible/NIR spectroscopy and validated by a hyperspectral imaging system with decision tree algorithm. The on-line testing results showed that the multispectral imaging technique can be used effectively for detecting feces (from duodenum, ceca, and colon) and ingesta on the surface of poultry carcasses with a processing speed of 140 birds per minute. This paper demonstrated both multispectral imaging hardware and real-time image processing software. For the software development, the Unified Modeling Language (UML) design approach was used for on-line application. The UML models included class, object, activity, sequence, and collaboration diagram. User interface model included 17 inputs and 6 outputs. A window-based real-time image processing software composed of 11 components, which represented class, architecture, and activity. Both hardware and software for a real-time fecal detection were tested at the pilot-scale poultry processing plant. The run-time of the software including online calibration was fast enough to inspect carcasses on-line with an industry requirement. Based on the preliminary test at the pilot-scale processing line, the system was able to acquire poultry images in real-time. According to the test results, the imaging system is reliable for the harsh environments and UML-based image processing software is flexible and easy to be updated when additional parameters are needed for in-plant trials. Mention of trade names or commercial products in this article is solely for the purpose of providing specific information and does not imply recommendation or endorsement by the U.S. Department of Agriculture.     

基于卷积神经网络的织物缝纫平整度客观评价

针对大多数评价织物缝纫平整度等级方法对实验条件均有较高的要求,且织物种类及环境等因素对实验结果均有较大影响的现状,提出利用卷积神经网络分析织物缝纫平整度等级的方法,以提高等级分类的准确率和效率。设计了一个基于残差卷积神经网络的织物缝纫平整度客观评价模型,该模型以1000个普通织物的缝纫图像作为训练样本输入,得到缝纫平整度的分类结果,所选织物包含10种常见服装面料品类(塔夫绸、塔斯隆、雪纺、顺纡绉、尼丝纺、麂皮绒、天丝斜纹、真丝缎面、平布、交织绸)。研究结果表明:经200个测试集样本的验证,该模型的评价准确率达96%,与智能化评价以及建立预测模型方法相比,利用卷积神经网络分析织物缝纫平整度等级的方法,具有较好的准确率,且具备获取样本图像流程简单、效率高的优势。     

Contamination of carcasses with Salmonella during poultry slaughter

Successively slaughtered poultry flocks were sampled for Salmonella to study the relationship between gastrointestinal colonization of the birds and contamination of the carcasses after slaughter. Samples from 56 broiler flocks and 16 spent layer and breeder flocks were collected in six slaughterhouses. Salmonella isolates were serotyped and further characterized by pulsed-field gel electrophoresis (PFGE). Although only 7 (13%) broiler flocks were colonized with Salmonella at slaughter, carcasses of 31 (55%) broiler flocks were contaminated after slaughter. Concerning the layer and breeder flocks, 11 (69%) flocks were colonized in the gastrointestinal tract, but after slaughter, carcasses of all flocks were contaminated. The Salmonella status determined at the farm did not always correlate to the status at slaughter. On the other hand, the slaughter of Salmonella-colonized flocks did not always result in the contamination of the carcasses with the same PFGE types isolated from the gastrointestinal tract. When only uncolonized flocks were slaughtered, the carcasses of flocks were on some occasions still contaminated with Salmonella. This indicates possible cross-contamination from the slaughter equipment or transport crates. These observations show that it is difficult to reach the benefits of logistic slaughter in commercial poultry slaughterhouses.     

EFFECTS OF TRISODIUM PHOSPHATE ON POPULATIONS OF SALMONELLA TYPHIMURIUM AND SALMONELLA ENTERITIDIS IN VITRO

This study determined the bactericidal activity of different concentrations of trisodium phosphate (TSP) against Salmonella typhimurium and Salmonella enteritidis in vitro. One milliliter of each TSP solution (0.5, 1, 1.5 and 2% w/v) was mixed with 1 mL of each bacterial culture and left to stand at room temperature (18  ±  1C) for 0, 15, 30, 45, 60 or 90 min before enumeration of the surviving pathogens.
The effects of 1.5 and 2% TSP were apparent on both bacterial species from 0 min and bacterial counts reached undetectable ( < 1.0  ×  10¹ cfu/mL) levels immediately. The 1% TSP solution effected close to 4.20–5.72 log reduction after 15 min. However, the reduction effect of 0.5% TSP was a function of treatment time, and some culturable bacteria remained even after 90 min of treatment. These results suggest that the bactericidal effect of TSP solution on these pathogens depends on its concentration and on the duration of treatment.

PRACTICAL APPLICATIONS

Poultry has been identified as one of the most important reservoirs of Salmonella typhimurium and Salmonella enteritidis in food chain. Efforts to eliminate or reduce Salmonella spp. and other pathogens attached to poultry carcasses during processing have been made for many years. In 1992, the U.S. Department of Agriculture approved the use of trisodium phosphate (TSP) as a processing aid to reduce Salmonella spp. on raw poultry carcasses. The aim of the present study was to determine the bactericidal effects of different concentrations of TSP on pure cultures of these pathogens in vitro .     

Prevalence and risk factors for Salmonella and Campylobacter spp. carcass contamination in broiler chickens slaughtered in Quebec, Canada

An observational study was conducted to estimate prevalence and risk factors for Salmonella and Campylobacter spp. carcass contamination in broiler chickens. Eighty-two lots were sampled in four slaughterhouses located in the province of Québec, Canada, over a 10-month period. Carcass contamination was evaluated by the carcass rinse technique for about 30 birds per lot. Exposure to potential risk factors was evaluated based on data from questionnaires, meteorology, and cecal cultures. Multivariable binomial negative regression models were used for risk factor analysis at the lot level. The prevalence of Salmonella-positive carcasses was 21.2% (95% confidence interval: 15.7 to 26.7%). Significant risk factors (P < 0.05) associated with a higher proportion of positive carcasses within lots were Salmonella-positive cecal culture, low rainfall during transportation to the slaughterhouse, temperature of > or = 0 degree C during transportation to the slaughterhouse, and a > or = 4-h waiting period in shipping crates before slaughtering. The prevalence of Campylobacter-positive carcasses was 35.8% (95% confidence interval: 27.1 to 44.5%). Lots containing birds with Campylobacter-positive cecal culture results, lots of birds that were slaughtered at the end of the week, and lots with at least 20% of birds with digestive contents detected in the jejunum at time of slaughtering had a significantly higher proportion (P < 0.05) of contaminated carcasses. These results support the importance of preharvest control measures implemented during rearing to reduce contamination of the final product. Weather during transportation to slaughter and the day of the week that birds were slaughtered also were associated with carcass contamination; further studies are needed to determine the underlying mechanisms by which these factors influence carcass contamination.     

Determination of lean meat in pig carcasses with the Autofom classification system

A fully automatic classification equipment called Autofom for use at pig abattoirs has been developed and tested. It measures fat- and meat depth in carcasses by use of ultrasound. Measurement with Autofom is very fast (measuring up to 1250 carcasses per hour). It is non-invasive and low maintenance costs can be expected due to lack of moving parts. The repeatability and reproducibility have been estimated. An equation for calculating lean meat percentage on the basis of Autofom measurements has been calculated. The accuracy of the equation, expressed by residual mean square error and by standard error of prediction, is 1.84 and 1.95, respectively. A functionality test shows that, when installed on-line at the abattoir, 99% of the carcasses can be classified by Autofom. The equipment will be a suitable alternative to automatic on-line determination of lean meat percentage in pig carcasses.     

Evaluation of Data Mining Strategies for Classification of Black Tea Based on Image-Based Features

In this study, a new procedure based on computer vision was developed for qualitative classification of black tea. Images of 240 samples from four different classes of black tea, including Orange Pekoe (OP), Flowery Orange Pekoe (FOP), Flowery Broken Orange Pekoe (FBOP), and Pekoe Dust One (PD-ONE), were acquired and processed using a computer vision system. Eighteen color features, 13 gray-image texture features, and 52 wavelet texture features were extracted and assessed. Two common heuristic feature selection methods: correlation-based feature selection (CFS) and principal component analysis (PCA), were used for selecting the most significant features. Seven of the primary features were selected by CFS as the most relevant ones, while PCA converted the original variables into 11 independent components. These final discriminatory vectors were evaluated by using four different classification methods including decision tree (DT), support vector machine (SVM), Bayesian network (BN), and artificial neural networks (ANN) to predict the qualitative category of tea samples. Among the studied classifiers, the ANN with 7–10–4 topology developed by CFS-selected features provided the best classifier with a classification rate of 96.25%. The other methods assayed provided slightly lower accuracies than ANN from 86.25% for BN till 87.50% for SVM and 88.75% for DT. In all the cases, the accuracy of the classifiers increased when using the CFS-selected features as input variables in front of PCA obtained ones. It can be concluded that image-based features are strong characterizing factors which can be effectively applied for tea quality evaluation.     

振动胁迫下双孢蘑菇高光谱成像品质检测

为快速有效识别双孢蘑菇轻微损伤,以不同振动时间后不同损伤程度的双孢蘑菇为研究对象,采集400～1 000 nm的完好无损、振动60 s和振动120 s双孢蘑菇的近红外高光谱图像,发现3 种类型的双孢蘑菇在450～750 nm的光谱曲线有明显差异。比较标准正态变量变换、SG（Savitzky-Golay）平滑和多元散射校正等预处理方法,确定SG平滑为最优预处理方法。并将处理后的数据采用连续投影算法和竞争性自适应重加权算法提取不同损伤程度的特征波段;基于灰度共生矩阵提取500 nm波长特征图像感兴趣区域的纹理特征,分别将光谱信息和纹理特征信息作为输入,建立偏最小二乘判别分析（partial least squares-discriminant analysis,PLS-DA）、BP（back propagation）神经网络和极限学习机损伤程度识别模型。结果表明,两种特征集建模,PLS-DA模型均表现出最好的识别效果,PLS-DA模型训练集和测试集平均识别准确率为93.33%、91.11%和88.89%、86.67%。最后基于光谱-纹理融合信息建立PLS-DA模型,训练集和测试集总体识别正确率分别为97.78%、95.56%。结果表明,光谱-纹理融合信息建模预测效果优于单一特征信息建立的判别模型。因此,采用高光谱融合信息建模可以提高不同损伤程度的双孢蘑菇检测精度,为双孢蘑菇贮藏、分类提供理论支撑。