Development Of Nondestructive Detection Device For Apple Crispness Based On The Fusion Of Spectroscopy And Machine Vision

China has the largest apple planting area and highest yield in the world.Apple is one of the main fruits in China,which has high nutritional,functional and economic values.In order to meet the requirements of consumers for apple quality,the research on various index of apple is endless,and all kinds of apple grading products are constantly pouring into the market.The nondestructive detection and grading technique is one of the key techniques to improve the quality and economic value in the process of apple industrialization.The indexes covered by the existing researches on apple texture can not be evaluated comprehensively.Crispness is an important index to evaluate apple flavor and taste,but there is no complete system to detect crispness nondestructively.Therefore,it is significant to design a device which can detect crispness nondestructively,which is of great significance to the nondestructive detection of apple crispness and the objective evaluation.This thesis takes“Fuji”?“Qinguan”?“Ruiyang”in Shaanxi as the research object,and the main research contents and conclusions are as follows:(1)Design of crispness nondestructive detecting device based on the fusion of optical fiber spectroscopy and machine vision.The hardware and software system of the device were designed.Through serial communication with camera and spectrometer,the collected apple image information and spectrum information were sent to raspberry pi for processing and modeling;the program was compiled with Python and Open CV Machine Vision Library.The apple spectrum image,camera image,processed digital image,crispness values and grades were displayed on the display screen.(2)Determination of the relationship between crispness and firmness and establishment of quantitative prediction model for apple crispness based on firmness.The results showed there was a significant positive correlation between firmness and crispness of Fuji,Qinguan and Ruiyang apple,and the correlation coefficients were 0.917,0.900 and 0.910,respectively.The determination coefficients of the quantitative prediction models for crispness of Fuji,Qinguan and Ruiyang apple were 0.8412,0.8105 and 0.8282,respectively.(3)Establishment the grading model based on firmness and objective criteria of sensory evaluation.Apple crispness was classified into five grades:firmness,firm crispness,crispness,soft crispness and softness.The optimal crispness value was 2.62-3.20 kg.By objectifying the sensory evaluation,the accuracy rate was higher than 0.88,achieving a higher detection accuracy.According to the predicted crispness and the actual measured crispness,the corresponding crispness grade was determined.The results showed that the classification accuracy of the three apple cultivars can reach 90%.(4)Establishment of quantitative prediction model for apple crispness quality via partial least square metho based on the full wavelength and effective wavelengths.The prediction set determination coefficient R~2 of quantitative prediction model for Fuji apple based on raw spectra and the prediction set determination coefficient R~2 of quantitative prediction model for Qinguan apple based on S-G smoothing was 0.9205 and 0.9587,respectively,and the root mean square error was 0.1669 kg and 0.1741 kg,respectively.In order to eliminate redundant information and facilitate the development of portable testing instruments,the continuous projection algorithm(SPA)and the load factor method(x-LW)were used to extract effective wavelengths.The effective wavelengths of Fuji apple and Qinguan apple were 450.51 nm,476.80 nm,677.75 nm,750.72 nm and 542.51 nm,544.79nm,676.96 nm,718.29 nm,respectively.Partial least square method was used to establish the quantitative prediction model based on effective wavelengths.Artificial neural network(ANN)was used to predict and verify the model.The results showed that the prediction accuracy of the quantitative prediction model for Fuji and Qinguan apple based on the above effective wavelengths were 0.9131 and 0.9641,respectively.It is feasible and reliable to develop the portable instrument by the above effective wavelengths and nondestructively detect.(5)Establishment of dynamic models for apple texture quality based on spectral information under different preservation conditions.The optimal retention period of apple crispness could be predicted according to the initial spectral information.The experimental results showed that the crispness of Fuji and Qinguan apple cultivars declined with time.The correlation coefficients of the kinetic models for Fuji apple under room and refrigeration temperature were 0.956 and 0.87,respectively.The optimal crispness retention times of Fuji apple under room and refrigeration temperature were 5 weeks and 8 weeks,respectively.The correlation coefficients of the kinetic models for Qinguan apple under room and refrigeration temperature were 0.9958 and 0.9514,respectively,and the optimal crispness retention times of Qinguan apple under room and refrigeration temperature was 4weeks and 7 weeks,respectively.(6)Establishment of a discrimination model for apple crispness based on the fusion of appearance image and spectral information to improve the accuracy of nondestructively detecting and reduce the influence of apple appearance on the internal quality discrimination.The quantitative prediction model between apple crispness and RGB components was established via partial least square method.The prediction accuracy of the model was 0.9785.The experimental results showed that the mean value of relative error before and after negative compensation of apple appearance decreased from 8.29%to 5.55%,and the mean value of relative error before and after positive compensation of apple appearance decreased from 8.52%to 6.44%.The standard error of relative error before and after negative compensation of apple appearance decreased from 5.62%to 4.44%;the standard error of relative error before and after positive compensation of apple appearance decreased from6.22%to 5.3%.This showed that the mean and standard errors of relative errors are reduced,and the accuracy of on-line detection of apple crispness classification was up to 90%,so the apple crispness prediction is more accurate and reliable after apple appearance compensation.After compensation,the relative error of predicted crispness was reduced,which suggests that it is effective and meaningful to detect and predict apple crispness after the fusion of spectral information and apple appearance.