On-line Nondestructive Prediction Of Pork Quality Using Image And Spectroscopy Information

As the most popular meat resource in China, pork quality deserves great attention of consumers and quality supervision. The quality of meat and meat product plays an important role in supporting the qualified daily life, nutrition ingestion and safety of dietary for the people. In the meat industry, to supply high-quality products consistently for consumers is the key of wining in the market competition. Conventional ways for the assessment of meat quality are human inspector or use instrumental test and chemical analysis. However, these methods are tedious, laborious and costly, destructive and/or time-consuming, and consequently impracticable for on-line application. Over the last few years, machine vision and near-infrared spectroscopy (usually spanning the visible and near-infrared rang, Vis/NIR) were widely used to predict sensory and physical-chemical characteristics of meat and meat products. As fast and nondestructive, and even noncontact methods to predict meat quality, they are considered as the most promising for on-line prediction.The purpose of the present study was focused on the on-line prediction of quality traits of intact Longissimus Dorsi muscle (MLD), which was taken from fresh pig carcass. Digital images and Vis/NIR spectra were scaned from 212 samples using a prototype, which was designed to on-line determinate the sensory and physical-chemical characteristics of meat. An automatic image processing program were developed to eliminat the background of the image, select MLD region, segment the muscle region and the intermuscular fat region. Then the feature information of image, which be supposed to sensory quality relative, was extracted to evaluate the sensory scors of pork color, color uniformity and marbling. Partial least squares discriminatory analysis (PLSDA) and support vector machine discriminatory analysis (SVMDA) was employed to construct the evaluation models. Discrete wavelet transform (DWT) was employed to eliminate noise. Partial least-squares regression (PLSR) based on the wavelet de-noised spectra combined with different spectra pretreatments were explored to predict the physical-chemical characteristics. In addition, several variable selection methodologies were applied for producing parsimonious calibration model relating the spectra to pork physical-chemical characteristics.The main contents and conclusion were:(1) A prototype integrated computer vision and spectrum analysis technology was developed for the purpose of pork quality determination on-line. RGB image and Vis/NIR spectral information of 212 MLD samples were obtained using the prototype at the speed of 0.25 m/s. The color and shape of the image presented an excellent reproducibility, and the spectrum was in accord with previous study of off-line scnned. However, due to the on-line scanned condition, the complex and variability properties of fresh pork, the diverse surface physical structure and other factors, the intensity of image was inhomogeneous, and a baseline shift was involved in the spectrum.(2) The evaluators graded the similar classification for the color well, but great confusion for color uniformity in the rank of nonuniformity, rather nonuniformity and uniformity, variant scores for marbling trace and scruple. The curve of TPA indicated that the compressive deformation of pork is firstly from plastic deformation changing into elastic deformation, than changing into plastic deformation again until compression failure. The optimum termination criterion of TPA is 50% compression ratio. Repetitive pH measurements at different positions in each sample provided reasonably high reproducibility. Shear force values of four cores from a sample showed significant variation due that the physical structure of muscle and its fiber arrangements are inhomogeneous and the cores are soft and easily deformed. Correlation analysis of each quality trait to others showed that there was no significant correlation between different traits.(3) An automatic image processing program was developed to segment MLD images into background, muscle and IMF. In this program, the KSW algorithm combined with morphology operations was employed in eliminating the surrounding fat of MLD. And K-means clustering algorithm was applied to mask out the IMF region with the intensity inhomogeneity corrected. Before intensity inhomogeneity corrected, partial lean pixels with high intensity were misjudged as fat pixes, while partial fat pixels with low intensity were marked as leanpixes. The intensity map, which obtained through two-dimensional B splines curve fitting, provided smart intensity inhomogeneity imformation but ignored the step-variance at the boundaries of IMF and muscular regions. And the image after corrected by this intensity map could be segmented correctly via K-means clustering algorithm. The result indicates that only 9 images were failed to stable and accurate segmentation because of serious specular reflections on the pork sample surface.(4) Color features, including mean and standard deviation of red, green, blue, hue, saturation, and value componentes of the segmented muscle area, and image features of the segmented IMF were extracted from segmented images. Both PLSDA and SVMDA models showed that discriminant accuracy ratios for color evaluation is about 70%, and greater than 77% for marbling assessment, but lower than 66% for color uniformity evaluation in prediction set.(5) The optimal combination of denoising was wavelet db6, decomposition level 6, and soft thresholding with minimaxi threshold estimator. After applying the first derivation, the prediction ability of the calibration models for all parameters improved. This means that the negative impact of the translational error independent of wavelength in the reflectance spectra which greatly affects the calibration was eliminated after the first derivation pre-processing. The model based on spectra that were denoised under the optimal DWT parameters and appropriate pretreatment gave reasonable performances. The prediction correlation coefficients of the models for the main chemical conponets (IMF, protein and moisture) were greater than 0.87 in calibration and prediction sets, and the ratio of prediction to deviation (RPD) in calibration and prediction sets (calibration/validation) were 2.61/1.72,2.37/1.74,2.45/2.01 respectively. The prediction correlation coefficients of the model for the pH values were greater than 0.9, and RPD (calibration/validation) were greater than 2. For hardness, springiness, cohesiveness and resilience, RPD were 1.89/1.33,1.97/1.82,1.71/1.60 and 1.83/1.63. Unfortunately, no useful model could be constructed for shear force. The method of variable importance in projection renovated the simpler models with a few important variables following slightly deteriorate.The results obtained in this study indicated that machine vision and Vis/NIR spectroscopy is a promising technique to simultaneously predict the sensory and physical-chemical quality attributes of intact fresh pork on-line. The study laid foundation for developing intelligent determination equipment of fresh pork quality for meat industry, and provided bases for establishing a system of raw meat classification with different applications.