Detection Of Endogenous Foreign Objects In Chinese Hickory By Spectral Imaging At Pixel Level

Chinese Hickory is a nut that is popular among consumers.It is approximately circular in shape and about half the length of American nuts.It is difficult to identify small and harmful shell fragments by visual inspection,because they are not only small in size,but also similar in color to Chinese Hickory flesh.Ingestion of shells can cause physical harm to consumers,so a method for effectively identifying small nuts is needed to provide a new method for the monitoring and sorting of intelligent production lines for Chinese Hickory kernels.Existing researches use objects' optical characteristics such as reflection and absorption of light and image features such as color and texture to identify foreign objects.This study comprehensively used spectral imaging technology(hyperspectral imaging and multi-spectral imaging),traditional machine learning and deep learning methods,with Chinese Hickory as the research object,and endogenous foreign body shell as the detection target.The research of pixel-level detection methods,the main research contents and conclusions are as follows:(1)The pixels of Chinese Hickory kernel(outer meat,inner meat)and Chinese Hickory shell(inner shell,outer shell)in hyperspectral images were extracted,and the spectral differences of outer meat,inner meat,outer shell and inner shell pixels were analyzed.The average spectrum and first-order differential spectrum of the above components show that: in the 400-900 nm band,the outer and inner meat,outer shell and inner shell have similar waveforms,and the inner meat spectral reflectance is higher than the other three components The spectral reflectance curve has a peak at 880 nm.The first two principal component scores of the spectrum of each component of hickory showed that there was no obvious boundary in the distribution area of each component.There was more overlap between the outer meat,outer shell and inner shell,and there was relatively less overlap between the inner meat and other components.(2)Compare the effects of hyperspectral imaging technology combined with traditional machine learning and deep learning methods on the detection results of endogenous foreign bodies.The results of model detection based on hyperspectral images show that: 1)Principal component analysis-K nearest neighbor(PCA-KNN)model overall discrimination accuracy rate is 94.0%,and support vector machine(SVM)model overall discrimination accuracy rate is 93.0%;2)The overall discrimination accuracy of the one-dimensional convolutional neural network(1D-CNN)model is93.7%.The overall discrimination of the two-dimensional convolutional neural network(2D-CNN)model and the convolutional neural network-long-short-term memory network(CNN-LSTM)model The accuracy rates are 98.5% and 99.0%respectively;3)Compared with the traditional machine learning model,the deep learning model has a stronger classification ability for Chinese Hickory spectral data as a whole,but the model training time is longer.(3)The multi-spectral imaging system for endogenous foreign body detection compares the effects of multi-spectral imaging technology combined with KNN and CNN-LSTM methods on the detection results of nut endogenous foreign bodies.Using the characteristic wavelength method based on genetic algorithm,the five characteristic wavelengths screened are: 462,536,671,743 and 887 nm.The results of model detection based on multispectral images show that: 1)The accuracy of the test set of the KNN model and CNN-LSTM model is 88.3% and 81.0%,respectively,and the KNN model is superior to the CNN-LSTM model;3)Although the model based on the multi-spectral image is slightly inferior in the accuracy of the test set,the multi-spectral image has the advantages of less band and small data volume.