Rapid And Non-destructive Detection Of Chicken Meat Quality Based On Hyperspectral Imaging Technique

Chicken meat plays an important role in people’s daily life as it can provide abundant proteins, vitamins and minerals. With the improvement of people’s life, consumers pay more attention to chicken meat quality and safety. Therefore, it is significant to detect chicken meat quality. However, traditional methods for quality detection are time-consuming, tedious and destructive, which cannot meet the requirements of on-line inspection in the modern meat industry. This study applied visible/near-infrared(400-1000 nm) hyperspectral imaging technique coupled with chemometric methods and image processing methods for detecting chicken meat quality attributes, including physical attributes(tenderness and color(L*, a*, b*)), chemical attributes(thiobarbituricacid reactive substances(TBARS) and hydroxyproline), and quality classification. The encouraging results can provide a possible way for the meat industry to monitor quality variations. Specifically, the main research contents and results are shown as follows:(1) Rapid and non-destructive prediction of chicken meat tenderness and color(L*, a*, b*) were achieved based on visible/near-infrared hyperspectral imaging technique. First, mean spectra of region of interests(ROIs) in hyperspectral images were extracted. Then, three traditional spectral pre-processing methods(multiple scatter correction(MSC), standard normalization variable(SNV) and Savitzky-Golay smoothing(SG)) were respectively used for spectral de-noising. Finally, the PLSR model built with MSC spectra showed the best results and therefore used as the optimal method for spectral pre-processing. In order to simply the PLSR model, two classical wavelength selection methods namely regression coefficients(RC) and successive projections algorithm(SPA) were respectively used. Based on the selected wavelengths, two optimized models namely RC-PLSR and SPA-PLSR were established, in which the SPA-PLSR models showed better performance as they had higher values of regression coefficient in prediction(Rp)(tenderness: Rp=0.740, RMSEP=11.147; L*: Rp=0.876, RMSEP=1.518; a*: Rp=0.897, RMSEP=0.882; b*: Rp=0.959, RMSEP=1.175). Finally, distribution maps of color(L*, a*, b*) were created by transferring the SPA-PLSR model to each pixel in some representative hyperspectral images.(2) The potential of visible/near-infrared hyperspectral imaging for predicting TBARS and hydroxyproline content in chicken meat was investigated. First, PLSR models for TBARS and hydroxyproline prediction were respectively established based on full spectra. Then, optimal wavelengths of TBARS and hydroxyproline were respectively selected using the RC method.Based on optimal wavelengths, three linear regression models namely RC-PLSR, RC-MLR, and RC-PCR were respectively built, in which the RC-MLR models showed the best performance in predicting TBARS(Rp=0.871, RMSEP=0.124) and hydroxyproline(Rp=0.903, RMSEP=0.036) content. Finally, an image processing algorithm was developed to create distribution maps of TBARS and hydroxyproline values in some representative chicken meat samples.(3) Combination of spectra and texture data of hyperspectral imaging for differentiating between free-range and broiler chicken meats were investigated. First, hyperspectral images of chicken meat samples were obtained, and mean spectra of ROIs in the hyperspectral images were extracted. Then, SPA was used to select optimal wavelengths. On the other hand, principal component analysis(PCA) were carried out in all hyperspectral images and consequently, the first and the second principal components(PC1 and PC2) images were determined as the optimal characteristic images. Then, gray-level gradient co-occurrence matrix(GLGCM) was implementd on PC1 and PC2 images to extract 30 textural variables in total. Based on the full spectra, optimal spectra, texture data and data fusion, least-squares support vector machine(LS-SVM) and artificial neural network(ANN) models were respectively built, in which the LS-SVM model based on data fusion gave the best results with high correct classification rate(CCR) of 95% for the prediction samples, demonstrating that combining spectra with texture data were effective for differentiating between free-range and broiler chicken meats.