| Lingwu long jujube,which is a geographical indication product of Ningxia,is sweet and sour,rich in soluble sugar,phenols and other nutrients,and is popular among consumers.However,long dates are susceptible to various injuries during picking,grading,transportation,and processing due to man-made,mechanical collisions,extrusion,and vibration.Bruising is one of the most important injuries and a major problem in reducing the added value of jujube,which can lead to flavor deterioration and increase the risk of microbial infestation,while not dealing with bruised fruit in a timely manner can also cause decay and deterioration of the surrounding intact jujube.The grading of jujube fruits with different bruise levels and low price treatment can further reduce the loss and improve the economic efficiency of jujube industry.Therefore,this paper takes Lingwu jujube as the research object and fully exploits the data acquired by the NIR-HSI system combined with deep learning algorithms to discriminate intact and different grades of bruised jujube to find a fast nondestructive,efficient and accurate discriminative method in order to provide a detection method for online detection of jujube bruises,and the main findings are as follows:(1)A study of bruise grade discrimination of Lingwu jujube by fusion of spectral and image texture information.The quantitative bruising experiments were carried out on Lingwu jujube using the bruising device,and four different bruising grades(Ⅰ,Ⅱ,Ⅲ and Ⅳ)of jujube samples were obtained.The hyperspectral images of intact and bruised jujubes were collected using a nearinfrared hyperspectral imaging system,and the ENVI 5.1 software was used to extract the region of interest and calculate the mean reflectance spectra The hyperspectral image data were sequentially masked and subjected to principal component analysis,and the images with the highest principal component contribution were selected and texture features were extracted using a grey-scale co-occurrence matrix.Various pre-processing algorithms were used to pre-process the spectral data,and PLS-DA and SVM discriminant models for intact and different types of bruised long dates were developed.The results showed that the De-trending-PLS-DA model worked best,and the accuracy of both the calibration and prediction sets of the model was 90%.The PLS-DA and SVM discriminant models were developed by different feature wavelength algorithms,and the results showed that the SPA-PLS-DA model was the most effective,with 90%accuracy in both calibration and prediction sets.The results show that the De-trending-SPA-COR-PLS-DA model works best,with 92%accuracy in both correction and prediction sets.It is demonstrated that the fusion of spectral and image texture parameters can improve the prediction performance of the model to a certain extent.(2)A study on bruise grade discrimination of Lingwu jujube by fusion of atlas depth features.Hyperspectral images of intact and bruised jujubes were collected using a near-infrared hyperspectral imaging system.The deep features of the spectral data,image data,and spectral and image data fused by data-level fusion were extracted using a stacked autoencoder network,and PLS-DA and SVM discriminative models were developed for intact and bruised jujubes respectively.The results showed that the PLS-DA model-based results were optimal,and after dimensionality reduction by SAE network,the dimensionality of single spectral data was reduced from 256 to 27 dimensions,and the accuracy of model correction set and prediction set were both 94%;single image data was reduced from 900 to 38 dimensions,and the accuracy of model correction set and prediction set were 72%and 68%respectively;the atlas fusion data was reduced from 1156 to 50 dimensions,and the accuracy of model The accuracy of the calibration set and prediction set was 95%and 94%,respectively.A single spectral data reflects more valid information related to bruised dates than image information when distinguishing between intact and different types of bruised dates,but the fusion of spectral and image data proved to be more effective.(3)Deep learning-based fusion of spectral and image data to discriminate bruised grades of Lingwu jujube.Hyperspectral image data of intact and bruised jujubes were collected using a near-infrared hyperspectral imaging system,and the spectral data and image feature information were fused at the data level and used as input to a convolutional neural network and a long shortterm memory network to finally establish a discriminatory model for intact and bruised jujubes.The results showed that the 1D-CNN model had 100%accuracy in the training set and 97.6%accuracy in the prediction set,while the LSTM model had 100%accuracy in the training set and 96%accuracy in the prediction set.The fusion of multi-source information of long date samples combined with the deep learning algorithm model helps to further improve the model performance and provides a reference for other agricultural products for rapid nondestructive testing. |
PDF Full Text Request