Study On Rapid And Non-destructive Detection Of Egg’s Quality By Acoustic Resonance, Machine Vision And NIR Spectroscopy Techniques

Automatic measurement and classification of egg’s quality is very significant in solving the food quality& safety which are consumer’s concern, and improving its market value and competition. In this work, the acoustic technique combined with near-infrared (NIR) spectroscopy and computer vision were employed to automatic measurement and classification of egg’s quality. This research was explored in non-destructive detection of external quality (stiffness and cracks) and internal quality (freshness) of chicken eggs. The main achievements are summarized as follows:1. Development of eggshell quality analyzing system. A system based on acoustic resonance, as the basic experiment, was developed to detect eggshell quality. It was achieved by analysis of measured frequency response of eggshell excited by a light mechanical. A software program was written in Labview that allows a fast acquisition and processing of the response signal. In the system, material and parameters were optimized to acquire more effective signals. Five features variables were extracted from frequency response signals, and linear discriminant analysis (LDA) model was used to discriminate intact eggs and cracked eggs. The identification rate of LDA model was 86.1%. The esperimental results show that it is feasible to use acoustic resonance system to analysze quality of eggs.2. Detection of eggshell cracks by acoustic resonance technique. Two self-adaptive filter methods (Normalized least mean square, NLMS, and recursive least squares, RLS) were used comparatively to process response signals. It was found that with RLS signal processing, signal-to-noise ratio was remarkably enhanced. Ten features variables were exacted from response frequency signals. Genetic algorithms (GA) and stepwise regression algorithms were used comparatively to select features variables. The performance of GA was better than that of stepwise regression. Four features variables extracted by GA were used as input vectors of discrimination model. Linear (LDA; K-nearest neighbors, KNN) and non-linear (Artificial Neural Network, ANN; Support Vector Machine, SVM) pattern recognaization approaches were used to build calibration models. The optimal model was obtained when SVM algorithm was used, with the identification rates of 95.1%in calibration set and 97.1%in prediction set, respectively.3. Measurement of eggshell stiffness by acoustic resonance technique. Acoustic response signals of eggs combined with Partial least squares (PLS) algorithms were used to build calibrating model of egg’s stiffness. Synergy interval PLS (si-PLS), genetic algorithm PLS (GA-PLS) and GA-siPLS algorithms were used comparatively to select features variables, and build models. The performance of the final model was evaluated according to root mean square error of prediction (RMSEP) and correlation coefficient (R) in prediction set. Experimental results showed that the GA-siPLS model got acceptable performance with the fewest frequency variables. The optimal model was achieved with R= 0.7591 and RMSEP= 3.55 in the prediction set.4. Development of on-line measurement system for eggshell quality. According to the basic research, an on-line system based on acoustic resonance combined with digital signal processing (DSP) system was developed for the measurement of eggshell quality. In this system, automatic impacting of egg can be achieved, and then the response signal is acquired, processed, and analyzed. The speeds in commercial eggshell measurement graded up to 5 eggs per second, which can meet the on-line detection. The identification rates of intact eggs and cracked eggs were 97%and 95.5%, respectively. The correclation coefficient (R) between the stiffness measured by references method and acoustic response signal was 0.704, and mean square error of prediction was 2.2.5. Measurement of egg’s freshness by machine vision technology. RGB image of eggs were collected based on transmission imaging, and B components of RGB image was exacted for further analysis.14 variables were exacted, and 7 features variables were finally chosen by stepwise regression. BP-ANN was used to build calibration model, and Haugh units was used as the index of freshness. The correlation coefficient (R) etween the pridiction value and the actual one was 0.675 in prediction set.6. Analysis of freshness of eggs by near infrared spectroscopy (NIR) technique. Principle component analysis (PCA) and independent component analysis (ICA) were employed to extract useful information from samples, and eliminate much overlapped information. Compared with PCA, ICA shows more effective in extracting useful NIR spectral information of eggs. Genetic algorithm neural network regression (GA-NNR) was performed to calibrate the regression model. Some parameters of the regression models were optimized by cross-validation. The number of ICA factors included in ICA-NNR was optimized. When 7 ICs were used, the optimal model was obtained with root mean square error of prediction RMSEP=2.433 and the correlation coefficient R=0.879 in the prediction set. Traditional PLS model was also used to build calibration model, and the approach of augmented partial residual plots (APaRPs) combined run test were employed to test the linear degree between NIR prediction and actual measurement. A non-linearity result was obtained, and GA-NNR got better performance than PLS model.7. One-class classification was attempt to solve the classification problem due to imbalance number of training samples in eggs quality detection. One-class support vector machine (OC-SVM) was performed to detect intact eggs and cracked eggs. Conventional models (LDA, SVM) got poor performance when the ratio of intact eggs and cracked eggs was 11:1, but the optimal OC-SVM model could be achieved with identification rates of 90%. Support vector data discribption (SVDD) was performed to detect fresh eggs and un-fresh eggs. Conventional models (PLS-DA, KNN, ANN, SVM) got poor performance when the ratio of fresh eggs and un-fresh eggs was 13:1, but the optimal SVDD model could be achieved with identification rates of 93.3%.This research offers new ideas for automatic measurement and classification of eggs quality, which were great significant in improving the level of rapidly and non-destructive evaluation of egg’s quality.