Surface Defects Detection Of Apricot By Hyperspectral Imaging

Apricot,as one of the important characteristic fruits in Xinjiang,is loved by consumers for its nutrient-rich and medicinal value.Apricot grading is the core of the post-harvesting treatment process,in which the apricot surface defects is one of the important indicators of grading.However,the current domestic products processing of apricot and apricot lacks of mature detection devices.And it mainly rely on artificial sorting with low efficiency,high intensity labor,and whatsmore the sorting effect is difficultly to control.Hyperspectral imaging technology integrate with spectral and image information.Therefore,hyperspectral imaging is used to detect for apricot surface defects.The paper,with fresh apricot in Xinjiang as the study objects,and this research aims to investigate the potential of using visible near infrared(Vis-NIR)hyperspectral imaging for detection of sound,worn,mildew,insect apricot.Therefore,the research will provide theoretical basis for the development of online apricot surface defects detection system using multispectral imaging.The main methods and results are as follows:(1)Hyperspectral image data acquisition parameters were identified,and apricot samples were imaged using a hyperspectral imaging system.In the hyperspectral image,region of interests(ROIs)were manually selected to extract original spectra of individual samples.(2)Savitsky-Golay(SG),Standard Normal Variate Correction(SNV),Multiplicative Scattering Correction(MSC)were used for the pretreatment of original spectra data.And then SVM was applied to set up prediction models of defection of apricot.The results showed that the model of SVM based the original spectra and SG can create the better prediction result,the accuracy of identifying the defects of apricot was up to 93.3%.(3)Principal component analysis(PCA)was performed on the extracted spectra to select four special bands(495nm,570 nm,729nm and 891nm)for defects detection.The SVM,PLS-DA and BP neural network model based on the full and special wavelengths had been compared.The results showed that the established SVM model based on special wavelengths was superior to the full wavelengths,in which linear Kernel of SVM based on special wavelengths identification accuracy was 100%.The PLS-DA model based on full wavelengths was better than the special wavelengths.The identification results of the BP neural network based on all wavelengths superior to the special wavelengths.In addition,the SVM model based on special wavelengths superior to the PLS-DA and BP neural network.(4)PCA analysis was conducted on the spectral images using the full and special wavelengths to generate PC images,respectively.Masking,threshold-based segmentation,and morphologic operations were applied on the generated images to identify defective goals on the apricot.The results showed that the accuracies of identifying the sound,worn,mildewand insect apricot were 100%,38.3%,100% and 88.3% in the PC images based on the full wavelengths respectively,and 100%,80%,100% and 95% respectively in the PC images based on the special wavelength.The overall identification accuracy based on the full and the special wavelengths were 81.7% and 93.8% respectively.(5)MNF was conducted on the spectral images using the full and the special wavelengths to generate images,respectively.The results showed that the accuracies of identifying the sound,worn,mildew,insect apricot were 38.3%,33.3%,53.3% and 50% in the MNF images based on the full wavelengths,and the special wavelengths obtained better identification accuracy of sound and insect apricot that develop to 73.3% and 71.7%,respectively.The overall identification accuracy based on the full and special wavelengths were 43.8% and 60%respectively.As a whole,the overall recognition results of the PCA superior MNF.PCA and MNF were conducted based on the special wavelengths obtained better overall identification accuracy of apricot that develop to 93.8% and 60% respectively.Therefore,PCA analysis based on the special wavelengths can be more effective in identifying defects apricot and normal apricot.(6)In order to improve the detection rate of apricot surface defects,double band ratio analysis was conducted on the unrecognized worn apricot spectral images using the special wavelengths to generate band-ratio images,and the 570nm/891 nm band ratio image was selected to detect.The results showed that the identification accuracy of worn apricot was increased from 80% to 88.3%.