Study On High-throughput Online Rapid Nondestructive Detection Of Egg Multiple Qualities

Detection of egg quality is a key section in commercial processing,which is of great significance to improve economic value of egg and quality of people's life.Especially high-throughput online detection has a positive effect to promote production automation level and industrial development of egg in our country.In order to achieve high-throughput online detection of egg quality,spectral analysis and machine vision technology were employed to detect multiple qualities(freshness,scattered,size,shape and broken)of egg in this study which was combined with actual production demand of egg.The main research contents and conclusion were summarized as follows:1)Rapid quantitative detection of egg freshness based on fiber spectrum technology.Transmittance spectra of eggs were collected by detection device established by oneself.Different models built by partial least squares regression and support vector regression combined with Savitzky-Golay smoothing filter,multiplicative scatter correction,standard normal variate,first derivative and second derivative were compared to find that the prediction effect of FD-SVR model was better and SVR models were superior to PLSR in general,indicated that implicit non-linear relationship between egg freshness and spectral information was extracted by SVR better.Linear dimensionality reduction principal component analysis and non-linear dimensionality reduction locally linear embedding were used to process spectrum data after FD for simplification of quantitative model,which achieved the fast detection of freshness.The result of model indicated that effective information was extracted better and dimension reduction effect was more obvious than PCA by LLE.The correlation coefficient and root mean square error of calibration were 0.922,7.21,and prediction were 0.911,8.80.The root mean square error of cross validation of LLE-SVR decreased 0.79.The study showed that it is feasible for the rapid quantitative detection of egg freshness based on fiber spectrum technology,and it provides a theoretical research for future high-throughput online detection of egg freshness.2)Online rapid identification of scattered egg based on fiber spectrum technology.Transmittance spectra of eggs were collected dynamically on single channel eggtransport mechanism that speed was 5000 eggs per hour.Successive projections algorithm and competitive adaptive reweighted sampling were used to select optimal wavelengths for different spectrum pretreatment data,which found that the characteristic wavelength number of SPA was in general less than CARS.Afterwards,multiple classifiers were built by partial least squares discriminatory analysis,classification and regression tree,k-nearest neighbor classification and soft independent modeling of class analogy combined with characteristic wavelengths.Five classifiers were chosen on the basis of number of variables and identification rate.Finally,SNV-SPA-PLS-DA model was applied to recognize scattered egg online by the comparison of detection time that the number of characteristic variables was 3,the detection time of a egg was 55.733 ms and the accurate rate was 97.14%.It provides a technological method for high-throughput online identification of scattered egg based on spectroscopy.3)Study on high-throughput online vision detection of egg size and shape.A group eggs image high-throughput online collecting system was designed for automatic acquisition of image.The communication between host computer and slave computer and the image collecting function were achieved by Visual C++ software,and trigger signal of photoelectric switch was received by STC89C52.Group eggs images were collected dynamically on six channels transport mechanism that speed was 30000 eggs per hour.The high-speed transmission effect to egg image was eliminated by pretreatment method that was less but effective.Outline of egg was reconstruct by least-squares ellipse fitting combined with convex hull algorithm for solving egg concave phenomenon because of light leak.The major and minor axis were revised according to analyze reason of distortion and a linear regression model between the number of pixels and its actual size was built.The correlation coefficient of combining with convex hull algorithm was more than direct least-squares ellipse fitting which was 95.66% and94.39%.The result shows that least-squares ellipse fitting combined with convex hull algorithm was better for extracting egg outline.84 eggs were selected for testing,and the grading accuracy of size and shape were 90.5% and 89.3%.The single egg only required52.762 ms to be detected.This method is believed to achieve high-throughput online detection of egg size and shape.4)Study on high-throughput online vision detection of scattered egg.Group eggs images were collected dynamically on three channels transport mechanism that speed was 15000 eggs per hour for improving the detection efficiency of scattered egg further.The target image was obtained through removing useless background by the method of egg size and shape detection.Color component average variables of RGB and HSV space were extracted to build models of random forest and partial least squares discriminatory analysis.It was found that the classifier effect of RGB and HSV uniting space color component was best and RF was better than PLS-DA.The accurate rate of RF model in RGB and HSV uniting space was 92.86% and the detection time of an egg was 127.4ms.It satisfies the high-throughput online requirement of 15000 eggs per hour and fulfills high-throughput online detection of scattered egg.5)Study on high-throughput online vision detection of broken egg.Group eggs images were collected dynamically on three channels transport mechanism that speed was 15000 eggs per hour.An egg should be decided whether broken according to comprehensive detection result of three images due to randomness of broken region location.The egg target image was obtained by effective pretreatment,besides Butterworth high-pass filtering and gray image enhancement were used to highlight broken characteristic,which spot noise region was emerged at the same time.Morphological features(circularity and length-width ratio of minimum enclosing rectangle)of different regions were extracted to build BP artificial neural networks model optimized by particle swarm optimization for distinguishing of broken and noise regions which recognition rate reached 99.44%.It indicated that the generalization ability and robustness of PSO-BP-ANN was better than BP-ANN.Spot noise area was removed and120 eggs were verified which recognition rate of broken and intact egg were 91.67% and95%.The single egg required 201.24 ms to be detected which satisfied the requirement of high-throughput online detection.