Yud-Ren Chen

Development of hyperspectral imaging technique for the detection of apple surface defects and contaminations

A high spatial resolution (0.5–1.0 mm) hyperspectral imaging system is presented as a tool for selecting better multispectral methods to detect defective and contaminated foods and agricultural products. Examples of direct linear or non-linear analysis of the spectral bands of hyperspectral images that resulted in more efficient multispectral imaging techniques are given. Various image analysis methods for the detection of defects and/or contaminations on the surfaces of Red Delicious, Golden Delicious, Gala, and Fuji apples are compared. Surface defects/contaminations studied include side rots, bruises, flyspecks, scabs and molds, fungal diseases (such as black pox), and soil contaminations. Differences in spectral responses within the 430–900 nm spectral range are analyzed using monochromatic images and second difference analysis methods for sorting wholesome and contaminated apples. An asymmetric second difference method using a chlorophyll absorption waveband at 685 nm and two bands in the near-infrared region is shown to provide excellent detection of the defective/contaminated portions of apples, independent of the apple color and cultivar. Simple and requiring less computation than other methods such as principal component analysis, the asymmetric second difference method can be easily implemented as a multispectral imaging technique. Published by Elsevier Ltd.

Hyperspectral imaging for safety inspection of food and agricultural products

Development of effective food inspection systems is critical in successful implementation of the hazard analysis and critical control points (HACCP) program. Hyperspectral imaging or imaging spectroscopy, which combines techniques of imaging and spectroscopy to acquire spatial and spectral information simultaneously, has great potential in food quality and safety inspection. This paper reviewed the basic principle and features of hyperspectral imaging and its hardware and software implementation. The potential areas of application for hyperspectral imaging in food quality and safety inspection were identified and its limitations were discussed. A hyperspectral imaging system developed for research in food quality and safety inspection was described. Experiments were performed to acquire hyperspectral images from four classes of poultry carcasses: normal, cadaver, septicemia, and tumor. Noticeable differences in the spectra of the relative reflectance and its second difference in the wavelengths between 430 nm and 900 nm were observed between wholesome and unwholesome carcasses. Differences among the three classes of unwholesome carcasses were also observed from their respective spectra. These results showed that hyperspectral imaging can be an effective tool for safety inspection of poultry carcasses.

Hyperspectral-Multispectral Line-Scan Imaging System for Automated Poultry Carcass Inspection Applications for Food Safety

A hyperspectral-multispectral line-scan imaging system was developed for differentiation of wholesome and systemically diseased chickens. In-plant testing was conducted for chickens on a commercial evisceration line moving at a speed of 70 birds per minute. Hyperspectral image data was acquired for a calibration data set of 543 wholesome and 64 systemically diseased birds and for a testing data set of 381 wholesome and 100 systemically diseased birds. The calibration data set was used to develop the parameters of the imaging system for conducting multispectral inspection based on fuzzy logic detection algorithms using selected key wavelengths. Using a threshold of 0.4 for fuzzy output decision values, multispectral classification was able to achieve 90.6% accuracy for wholesome birds and 93.8% accuracy for systemically diseased birds in the calibration data set and 97.6% accuracy for wholesome birds and 96.0% accuracy for systemically diseased birds in the testing data set. By adjusting the classification threshold, 100% accuracy was achieved for systemically diseased birds with a decrease in accuracy for wholesome birds to 88.7%. This adjustment shows that the system can be feasibly adapted as needed for implementation for specific purposes, such as paw harvesting operations or prescreening for food safety inspection. This line-scan imaging system is ideal for directly implementing multispectral classification methods developed from hyperspectral image analysis.

On-line inspection of poultry carcasses by a dual-camera system

Abstract The Instrumentation and Sensing Laboratory (ISL) has developed a multi-spectral imaging system for on-line inspection of poultry carcasses. The ISL design is based on two principles: (1) wholesome and unwholesome birds have different chemical compositions of tissues and may have different skin color, and (2) unwholesome carcasses may have physical abnormalities which can be detected by computerized imaging. On-line trials of the multi-spectral chicken carcass inspection system were conducted during a 14-day period in a poultry-processing plant in New Holland, Pennsylvania, where spectral images of 13,132 wholesome and 1459 unwholesome chicken carcasses were measured. For off-line model development, the accuracies for classification of wholesome and unwholesome carcasses were 95% and 88%. On-line testing of the neural network classification models with combination of the filter information was performed. The inspection system gave accuracies of 94% and 87% for wholesome and unwholesome carcasses, respectively. This accuracy was consistent with the results obtained previously on laboratory studies. Thus, the inspection system shows promise for separation of unwholesome chicken carcasses from wholesome carcasses in poultry processing lines.

Automated detection of fecal contamination of apples by multispectral laser-induced fluorescence imaging.

Animal feces are a suspected source of contamination of apples by disease-causing organisms such as Echerichia coli O157. Laser-induced fluorescence was used to detect different amounts of feces from dairy cows, deer, and a dairy pasture applied to Red Delicious apples. One day after application, detection for 1:2 and 1:20 dilutions was nearly 100%, and for 1:200 dilutions (<15 ng of dry matter) detection was >80%. Detection after apples had been washed and brushed was lowest for pasture feces; detection for 1:2, 1:20, and 1:200 dilutions of feces was 100%, 30%, and 0%, respectively. This technology may encourage development of commercial systems for detecting fecal contamination of apples.

Neural network classification of wheat using single kernel near-infrared transmittance spectra

To investigate an accurate, rapid, and nondestructive method for wheat classification in inspection terminals, backpropagation neural network models were developed, based on single wheat kernel near-infrared transmittance spectra. Six classes of wheat were studied. Neural network models were optimized for two-class and six-class classification. The wavelength range of the spectra was 850 to 1049 nm. For two-class models with 200 input nodes, the average classification accuracy was 97% to 100%. For the six-class model with 200 input nodes, the average accuracy was 94.7%. The classification between hard red winter (HRW) and hard red spring (HRS) was least accurate among the six classes. For rapid classification, a narrower wavelength range, 899 to 1049 nm, with an interval of 2 nm, was proposed and shown to have little loss in accuracy. The most time-consuming two-class (HRW-HRS) model could be calibrated and validated in less than 7 mm. Prediction for new data was nearly instantaneous. A backpropagation neural network model with a learning coefficient of 0.6 to 0.65 and momentum of 0.4 to 0.45, without a hidden layer, was effective for wheat classification.

Classifying diseased poultry carcasses by visible and near-IR reflectance spectroscopy

Five classification techniques were compared for their accuracy in classifying normal, septicemic, and cadaver chicken carcasses, based on their optical reflectance spectra in the visible and near-infrared regions (504 - 888 nm). The techniques compared were the multiple- linear-regression, closest-class-mean, k-nearest-neighbor, artificial-neural-network (ANN), and principal-component/Mahalanobis-distance methods. The spectra were obtained with a diode array spectrophotometer system. The collection of the data and the development of the multiple linear regression model were described previously (Chen and Massie, 1992). The best results were obtained with the ANN model using the reflectances at the 8 optimal wavelengths identified by the multiple-linear regression method. The overall classification accuracy of this model was 91.6%. However, another ANN model with 192 inputs, which resulted in an overall accuracy of 90.4%, was preferred, because it utilized a broader range of reflectances (512.9 to 851.6 nm) without performing a wavelength search. This model yielded a 94.4% accuracy for the normal carcasses, 83.3% for the septicemic carcasses, and 94.3% for the cadaver carcasses.

Multispectral imaging for detecting contamination in poultry carcasses

A multispectral imaging system with selected optical filters of 542 and 700-nm wavelength was shown feasible for detecting contaminated poultry carcasses with high accuracy. The analysis of textural features based on co-occurrence matrix (COM) was conducted to determine the performance of multispectral image analyses in discriminating unwholesome poultry carcasses from wholesome carcasses. The variance, sum average, sum variance, and sum entropy of COM were the most significant texture features (P less than 0.0005) to identify unwholesome poultry carcasses. The feature values of angular second moment, variance, sum average, sum variance, and sum entropy did not vary with the distances and directions of COM for the spectral images. When a direction was equal to 0 degrees, the contrast was lower and the inverse difference moment and difference variance were higher (P less than 0.01) than any other direction in the visible spectral images. The characteristics of variance and sum variance texture feature of spectral images varied with the wavelength of spectral images and unwholesomeness of poultry carcasses as well. The sum variance of wholesome was higher (P less than 0.005) than unwholesome carcasses at the spectral image of 542-nm. The linear discriminant model was able to identify wholesome carcasses with classification accuracy of 83.9 percent and the unwholesome carcasses could be identified by quadratic model with 97.1 percent accuracy when textural features of spectral image at 700-nm wavelength were used as input data for models.

Online inspection of poultry carcasses using a visible/near-infrared spectrophotometer

The Instrumentation and Sensing Laboratory (ISL) has developed an industrial prototype diode-array visible/near-infrared (Vis/NIR) spectrophotometer system for inspecting poultry for diseased and defective carcasses on-line. The ISL design is based on the principle that wholesome and diseased and defective birds have different chemical compositions of tissues and may have different skin color. This visible/near- infrared spectrophotometer system has been tested off-line at 60 and 90 birds per minute. On-line trials of the visible/near-infrared chicken carcass inspection system were conducted during an 8-day period in a slaughter plant in New Holland, Pennsylvania, where 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 on 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.

Integration of visible/NIR spectroscopy and multispectral imaging for poultry carcass inspection

An integrated system which consisted of a visible/near-infrared spectroscopic subsystem and an intensified multispectral imaging subsystem was tested for its accuracy in separating abnormal (unwholesome) from normal poultry carcasses. The spectroscopic subsystem measured reflectance spectra of the poultry carcasses at wavelengths from 471 to 965 nm. For the multispectral imaging subsystem, the gray-level intensity of whole carcasses was measured using six different optical filters of 542, 571, 641, 700, 720, and 847 nm wavelengths. The preliminary results showed that, with the integrated system, there were no abnormal carcasses being misclassified as normal carcasses. When individual subsystem was used for classification, the error of the spectroscopic subsystem was 2.6% and that of the multispectral imaging subsystem was 3.9%. Thus, the integrated system could be used for separating carcasses into normal and abnormal streams. With perfect selection of normal carcasses in the normal carcass stream, the inspector needs only inspect the abnormal carcass stream. Thus, the through-put of carcasses of the processing line per inspector could be greatly increased.