Nondestructive Testing Of Fruits Based On Mosaic Spectral Imaging

With the development of agricultural technology,there are more and more fruit varieties on the market,and people have higher and higher requirements on the quality of fruit,so they hope to use some testing technologies to identify the external quality of fruit that cannot be intuitively judged by human eyes.Due to its wide band range,high spectral resolution and no damage to the tested object,hyperspectral imaging technology has become an important means of nondestructive testing of fruit quality.However,since most of the hyperspectral acquisition of fruits is based on line scanning,the spectrum acquisition device is bulky,the spectrum acquisition speed is slow,and it is easy to be subject to the rapid movement of platforms and objects.These factors make high-precision fruit nondestructive testing and sorting methods such as hyperspectral imaging mostly exist in the laboratory,which greatly influences the promotion of hyperspectral imaging in civilian daily testing and industrial application.This project based on this starting point,to build a set of mosaic spectral imaging fruit nondestructive testing system,using its acquisition device portable,obtain fruit spectral very fast and stable advantage,overcome the disadvantage of its low spectral resolution,studies the corresponding to terms with the spectrum of the mosaic to meet the high precision fruit nondestructive testing industry the purpose of fast separation.The main research work of this paper includes two aspects: the mosaic spectral imaging technology was used to identify the pesticide residues on the surface of cherry tomatoes and the dry jujube varieties were identified and classified.The main research results and conclusions are as follows:(1)A spectral imaging processing method suitable for mosaics--region of interest extraction and channel normalization is proposed.In this way,not only can spectral information of the entire sample surface be obtained completely,but also the complexity of spectral data can be reduced to a large extent and the fitting speed and accuracy of the model can be improved.(2)Five identification models of pesticide residues on tomatoes surface were established by using machine learning algorithm based on channel normalized spectrum,among which random forest(RF),gradient elevator(GBM)and support vector machine(SVM)identification models all obtained good results,with the identification accuracy reaching90.6%.The SVM model can ensure the accuracy while the fitting speed is only 8ms,while the GBM model has better robustness and identification performance,and its good identification accuracy and fitting speed make tomato pesticide detection expected to enter into daily life.(3)Based on the similar region of interest extraction and channel normalization,the identification model of 4 dried jujube varieties was established by cross-validation grid searching for the best parameters.The prediction accuracy of RF and GBM models for training sets can reach 100%,and that of test sets can reach 94% and 90% respectively.Due to the difficulty in exploiting the advantages of MLP models in small data sets,the identification accuracy is only 72%.The RBF-SVM(Gamma =1000,C=10000)model showed excellent performance in the dry jujube set,with the identification accuracy up to 98%.The cross-validation degree is 99%,which also shows that the model has strong generalization ability after the training of dry date data set.