Study On Determination Of Pork Quality By Near Infrared (NIR) Spectroscopy And Computer Vision

China is the biggest country of pork production and consumption in the world. With the improvement of the level of people's living, requirements for pork quality are higher and higher. At present, the regular detecting methods for pork quality are sensory evaluation and chemical analysis, which are much subjectivity, wasting time and expensive. This study attempted the feasibility on detecting pork quality by near-infrared (NIR) spectroscopy and computer vision. The main achievements of this work are summarized as follows:1.NIR spectroscopy was applied to building the quantitative model to quantitatively analyze the indexes of pork quality. The raw spectra of pork were preprocessed, and the calibration models of pork freshness and tenderness were built by classical partial least squares (PLS). The value of root mean square error of cross-validation (RMSECV) and correlation coefficient R(c) are 5.01 and 0.6758, 19.3 and 0.7187 in calibration set; the root mean square error of prediction (RMSEP) and correlation coefficient R(p) are 5.76 and 0.6982,24.9 and 0.4876 in prediction set. PLS is performed on full spectral region to calibrate global model, so some noisy spectral information inevitably weaken the performance of model. Therefore, In order to improve its precision and robustness, interval partial least-squares (iPLS), backward interval partial least-squares (biPLS), genetic algorithm partial least-squares (GA-PLS) and synergy interval partial least-squares (siPLS) were applied to selecting the efficient spectral regions for model of pork freshness. Experimental results showed that the four methods were able to produce better prediction models in relation to the full-spectrum model and the performance of siPLS model was the best in contrast to other three models. The siPLS model was achieved with R(c)=0.8332, RMSECV=3.75 in calibration set and R(p)=0.8238, RMSEP=4.17 in prediction set.2. The feasibility of the study were attempted to identify pork freshness with computer vision. In this experiment, pork muscles stored in a refrigerator at 4℃for different days were used for study target. The pork image was captured by CCD camera.12 color feature variables were extracted in RGB and HSI color spaces after pretreating the image. Next, the scores vectors were used to the input as pattern recognition after principal component analysis (PCA). Finally, the linear pattern recognition method which was linear discriminant analysis (LDA) was used to build identification model. When the principal component factor equaled to 10, the LDA model were the best. The identification rate equaled to 80% in training set and 80.48% in validation set. The experimental results showed that different freshness grades could be successfully identified through extracting suitable features variables.3.The feasibility of the study were attempted to identify pork tenderness with computer vision. In this experiment, pork was used for study target, and its tenderness grades were determined by TA-XT2i (Warner-Bratzler) and subjective judgment, which were tender, medium and tough. The pork image was captured by CCD camera. Four texture features (i.e., energy, equality, inertia and entropy) based on co-occurrence matrix were extracted from four directions (δ=0°,45°, 90°,135°), thus 16 feature variables were extracted. Next, the scores vectors were used for the input as pattern recognition after principal component analysis (PCA). Finally, KNN, the common linear pattern recognition method, was used to build identifying model. When the principal component factor equaled to 9, the KNN model was the best. The identification rate equaled to 89.13% in training set and 86.96% in validation set. The experimental results showed that different tenderness grades could be successfully identified through extracting suitable features variables.4. Multi-sensors fusion information of NIR spectroscopy and computer vision was applied to determine the pork freshness. The model based on two sensors fusion was built to discriminate pork freshness quality by BP-ANN. The fusion model is the best when the principal components factors from imaging information and NIR spectral information equaled to 6 and 12. The discriminating rate equaled to 100% in calibration set, and 97.22% in prediction set. The overall results showed that it was feasible to discriminate pork freshness with two sensors information from NIR spectroscopy and computer vision. The discriminating rate and robustness of the model from two sensors information fusion were better than the model built by single sensor.5.Hyper-spectral imaging technology was investigated to detect pork tenderness. Hyper-spectral images were collected between 408 to 1117 nm. The maximum shear force of these samples was obtained using the Warner-Bratzler attachment, and subjective judgment for the tenderness grade of pork was studied. The principal component analysis (PCA) was performed to select three optimal bands images. Then, four texture features (i.e., contrast, correlation, energy and homogeneity) based on co-occurrence matrix were extracted from each optimal band image, thus 12 variables for each pork sample in all. Finally, PCA was performed again to compress 12 features variables, and 10 principal components (PCs) were extracted as the input of back propagation (BP) neural net. Experimental results showed that discriminating rate was 96.15% in the calibration set and 80.77% in the prediction set. This study sufficiently demonstrates that the hyper-spectral imaging technology can be used to estimate tenderness grade of pork.This research offers a new idea to rapidly and nondestructively detect pork quality. It also has important significance in improvement of competition from our country in international market.