Research On The Appearance And Internal Multi Index Detection Of Apple Based On Hyperspectral Imaging

China is the largest country for apple planting and the yield has been ranked the first for nearly 10 years.Apples need to be graded precisely to suit the needs of different people.It is essential to establish the pipeline with high precision,fast speed and complete indexes to realize intelligent grading.It is of great significance for improving grading efficiency,saving labor costs and ensuring the competitiveness to imported varieties.In this paper,we took the Hanfu apple as the object of study.The external and internal indexes of apples were detected by the characteristics of integration of image and spectrum in hyperspectral imaging system.According to the index characteristics,we developed the targeted detection techniques and methods.A multi-index comprehensive grading simulation system was built through the common information in the detection.The main contents of this paper are as follows.(1)Analyzed apple index characteristics and determined the detection process according to the importance of the indexes.External indexes included size,shape,disease,and internal indexes included brix and firmness.Color validated grading results of the above indexes reference to the spectrum information as an auxiliary index.The order was established of size,shape,disease,brix,finnness and color to achieve comprehensive evaluation of external and internal quality.(2)The external detection was based on the information of 765nm and 904nm wavelengths and the detected index were size,shape and disease.In image preprocessing,median filter method was used to enhance 904nm image to detect size and shape and mask image of 765nm was used to detect disease.In the detection process,various indexes developed corresponding description methods and mathematical methods.MER moment method and pixel statistics method were used to detect the size,and the accuracy was 98.75%.The shape index,eccentricity and symmetry were described by the relevant information of minimum circumscribed circle and maximum inscribed of the apple contour.The shape index was described by the ratio of circumcircle area to incircle area.The eccentricity was described by the limit distance ratio of two circles.The symmetry was described by the ratio of segmented area of the line connected the tangent points.Apple shape was described from 3 angles with the comprehensive detection accuracy rate reached 95%.In disease detection,an improved manifold distance algorithm was proposed based on the difference of spectrum reflectance between disease area and normal area.By comparing L value of manifold distance,characteristic wavelengths suitable for disease detection(700nm,765nm and 904nm)were developed.According to the combination of spectrum information at different wavelengths,a BP neural network detection model was established.It was found that the reflectance spectrum of 765nm combined with 904nm could better detect the disease characteristics,and the accuracy was 96.25%.(3)In the internal index detection,the spectrum information of 543nm and 674nm wavelengths were used to detect apple brix and firmness at the same time.In the premise of double-sided image acquisition,the spectral reflectance waveforms of the region of interest(ROIs)with similar brightness were obtained.The waveform were smoothed by two order derivative combined with standard normal variable(SD+SNV).According to the test results of brix and firmness of ROIs,the characteristic wavelengths of two indexes were extracted by successive projections algorithm(SPA).Combined the distribution of characteristics wavelengths,two times SPA was proposed by swapping output values to search sharing information of two indexes detection.According to the information of sharing wavelengths in different environments,the prediction effects of least squares support vector machine(LS-SVM)and genetic algorithm(GA-BP)were compared.It was found that the best results was obtained from the optimal wavelengths of double-sided sampling(543nm and 674nm)by GA-BP,with the prediction correlation coefficient of brix(R)= 0.8476 and the mean square error(MSE)= 3.32,and of firmness with R = 0.7938 and MSE = 9.6.Besides,apple color was analyzed by the origin reflectance of 543nm and 674nm wavelengths.The ratio of limit difference algorithm and the ratio of color concentration of red to green were proposed.Apple color could be detected quantitatively.(4)A comprehensive online grading simulation system was designed by using MATLAB2013a.Detection program was written in the order size,shape,disease,brix and firmness.The online grading was simulated by identifying the apple image,spectrum and index information.Index database in different environments of the system was built reference to grading standard.According to the quantitative criteria,apples could be divided into Grade Super,Grade A,Grade B and substandard.In the end of the program,an optional verification link was added.The accuracy of grading results was verified by color detection.Simulation system interface was designed included user management,environmental parameters,index parameters,data display and other functions.Operators could rational use grading system by help information and comprehensive detection of apple multi-index was realized.