Research On Comprehensive Fruit Grading Method Based On Multi-source Information Fusion

China’s fruit production has steadily ranked first in the world for many years,but it is far below the position of developed countries in Europe and America in the world fruit export pattern.The main reason is that there is a gap in the fruit production and processing technology between China and developed countries.Therefore,it is necessary for China’s fruit industry to improve the level of automatic fruit grading.However,the existing fruit grading machines are so expensive that they are only suitable for processing large quantities of fruits.Meanwhile,the current research on automated fruit grading has problems such as single grading index,low grading accuracy and lack of comprehensive grading methods.Therefore,this thesis takes apples as an example to carry out the research of a comprehensive fruit rating method based on multisource information fusion which is suitable for small-scale fruit farmers and merchants.The details of the research are as follows:Regarding the detection of internal spectral characteristics of apples,this paper proposes the use of an integrating sphere to improve light collection efficiency,it also suggests using a competitive adaptive reweighting algorithm instead of the mainstream principal component analysis to realize the dimensionality reduction of spectral data.Meanwhile,multiple spectral data preprocessing methods are compared to select the optimal spectral preprocessing method to improve the accuracy of sweetness prediction.Ultimately,this paper combines the spectral processing methods of standard normal variate transformation,competitive adaptive reweighted algorithm,and standard normal variate transformation(SNV+CARS+SNV)with partial least squares regression to establish a sweetness prediction model for apples,with a correlation coefficient of 0.961 and a root mean square error of 0.313%.In terms of external feature detection of apples,this paper uses two methods,residual neural networks and traditional image processing,to extract features such as the size,color,and presence of defects of apples.Traditional image processing can solve the problem of difficulty in obtaining external feature labels during the process of building a neural network.Residual neural networks can effectively solve the problem of large errors caused by fixed parameters in the process of fruit automatic grading using traditional image processing methods.In the end,the accuracy of traditional image processing techniques for determining the presence of defects on the surface of fruits is 93.4%.The ResNet50 predicts the pixel values of fruit contour with a mean square error of 7.1 pixel and a correlation coefficient of 0.989.The ResNet34 predicts the color proportion of fruits with a mean square error of 4.61%and a correlation coefficient of 0.996.Based on the aforementioned algorithms,this paper independently designs a small-scale fruit rating detection platform,establishes a set of fruit comprehensive grading system using fuzzy mathematics principles,and develops the corresponding software.The software can detect the internal and external features of fruits in real time and perform comprehensive grading of the fruits.The research lays the foundation for the development of fruit grading equipment applicable to small-scale fruit farmers and merchants,which has a promising development prospect and provides technical support for the study of automatic grading of other fruit products.