| Peanuts are rich in nutrients and are an essential economic crop,as well as one of the major oil crops.In 2019,China’s total peanut production was 17.52 million tons,accounting for 37% of the world’s total production.Peanuts are highly prone to mold and produce carcinogenic aflatoxin due to improper conditions during storage and transportation.Therefore,it is essential to ensure the quality and safety of peanuts.Although traditional wet chemical methods can provide accurate detection results,these methods are costly,time-consuming and destructive,require highly skilled operators and accurate control of experimental conditions that are not suitable for automation or online monitoring.Therefore,there is an urgent need for a technology that can detect moldy peanuts in real time and non-destructive online before the production of products to meet the actual industrial production needs.Hyperspectral imaging technology can simultaneously obtain the spectral and spatial information of the target,the spectral information can be used to diagnose whether peanut is moldy,while the spatial information can help determine whether the kernel is moldy and mark the spatial location of moldy peanut kernel,thus enabling rapid and non-destructive automatic detection of moldy peanuts.In this study,three varieties of peanuts,namely Huayu,Silihong and Xiaobaisha,were placed in a constant temperature and humidity incubator to make them naturally moldy,and two gradients of moldy peanut samples were randomly sampled from the incubator on day 20 and 30,respectively.The hyperspectral images of three varieties of moldy and healthy peanuts were acquired using a hyperspectral imaging system to conduct the following research around the theory and methods of moldy peanut identification:1.Effect of irregular topography on hyperspectral images and correction.When collecting hyperspectral data,the topography of peanut particles will cause changes in peanut surface curvature thus leading to scattering effects,which leads to significant changes and differences in the peanut spectral curves of different location in the hyperspectral images,which in turn affects the subsequent identification of moldy peanuts.In this study,different levels of spectral intensity and spectral absorption features of peanut spectra were obtained by singular spectrum analysis(SSA),in which the first singular vector and singular value mainly contained baseline offset information,and the rest of the corresponding singular vectors and singular values contained multiplicative effect information;Subsequently,the effects of baseline shifts and multiplicative effects caused by topography were traced and analyzed by classification gradients,and a SSA-based correction model(CMSSA)was proposed for the correction of spectral data from different locations of peanuts;the results showed that CMSSA could eliminate the spectral variations caused by irregular topography of peanuts while preserving the chemical differences of interest,improving the accuracy of subsequent moldy peanut identification applications.2.Identification of moldy peanuts by combining hyperspectral images and spectral information.Three moldy peanuts identification models based on spectral information were developed using the collected peanut hyperspectral data,and compared and analyzed with each other.1)Moldy peanuts identification model based on key wavelengths: after eliminating the influence of topography using CMSSA,The optimal band combination was selected using genetic algorithm-continuous projection algorithm,and the moldy peanuts were identified using ensemble classifiers.The results show that the pixel-wise classification accuracies of the key wavelength-based ensemble classifier model were 92.17% and 89.00% for the training and test images,respectively,and their corresponding kernel-scale classification accuracies were 82.91% and 76.24%,respectively.2)Moldy peanuts identification model based on one-dimensional convolutional neural network(1D CNN): In this study,a 1D CNN including four convolutional blocks and two fully concatenated layers was developed to extract deep spectral features for identification of moldy peanut.First,the number of convolutional kernels of 1D CNN was optimized.Too many or too few convolutional kernels will reduce the performance of 1D CNN,and the model performance reached the optimum when its number was 15.By comparing the original spectral data and the CMSSA preprocessed spectral data respectively,it was found that the 1D CNN model based on the original spectral data outperformed the CMSSA pre-processed model.The pixel-wise classification accuracies of 1D CNN in training and test images were 97.28% and94.33%,respectively,and their corresponding kernel-scale classification accuracies were 90.98% and 86.48%,respectively,which were better than the model established by key wavelengths,indicating that the extraction of deep-level features by 1D CNN could improve the identification accuracy of moldy peanuts.3)Moldy peanuts identification based on spectral data augmentation: the spectral intensity feature space was constructed based on the singular values obtained by SSA,and a physical factor-driven spectral data augmentation method,local random mean(LRM),was proposed;spectral data augmentation based on LRM was performed under different number of samples and validated in combination with 1D CNN;the results showed that LRM could improve the classification accuracy of training and test images by 0.03～1.00% and0.13～1.16% respectively under different numbers of training samples,which was close to the performance of the model with increasing the same number of real samples.The overall results showed that 1D CNN combined with LRM can effectively identify moldy peanuts.Overall results indicated that hyperspectral techniques can be used for real-time,nondestructive detection of moldy peanuts;kernel-scale classification results can directly provide reference for online mechanical rejection of moldy peanuts;compared with identification models based on key wavelengths,1D CNN could more effectively use spectral information to identify moldy peanuts;based on 1D CNN,LRM can be used to identify moldy peanuts under different sample sizes.3.Identification of moldy peanuts based on hyperspectral images with spectralspatial information.Due to the changes in the acquisition environment and measurement conditions,the spectra of peanut samples are prone to the phenomena of "same spectrum for different objects" and "different spectrum for the same object",and it is difficult to obtain satisfactory classification results only by spectral information.Three moldy peanuts identification models based on spectral-spatial information were investigated.1)Moldy peanuts identification model based on fusing spectral and texture information: after eliminating the influence of topography using CMSSA,Seven key wavelengths sensitive to mold information were selected using a competitive adaptive reweighting algorithm,and the texture features of the key wavelengths were calculated using grayscale co-occurrence matrix;The results showed that the pixel-wise classification accuracy of the EC model fusing spectral and texture was 94.62% and92.24% for the training and test images,respectively,and their corresponding kernelscale classification accuracies were 84.63% and 78.24%,respectively,which outperformed the model using only spectral features.2)Moldy peanuts identification model based on multiscale point-centered convolutional neural network(MPCNN): the spectral and spatial features of the central pixel are automatically hierarchically represented by surrounding pixels to extract deep spectral-spatial fusion information,and an MPCNN model with patch as input was developed to identify moldy peanuts.Firstly,the hyperparameters such as the number of convolutional kernels and the input window size were optimized,and it was found that the optimal number of convolutional kernels and the optimal window size for moldy peanut recognition were 5 and 7×7,respectively.Second,the comparison of MPCNN models with CMSSA pre-processed data and the original data as input showed that CMSSA could enhance the spatial information related to peanut mold and thus improve the performance of MPCNN model.CMSSA-MPCNN achieves a classification accuracy of 99.69% and 98.43% at the pixel level for training and test images,respectively,and their corresponding kernelscale classification accuracies were 99.92% and 97.50%,respectively,outperforming1 D CNN and EC models fusing spectral and texture features.3)Moldy peanuts identification model based on point-centered convolutional neural network framework combined with embedded feature selection: When developing the CMSSA-MPCNN,some bands did not carry classification-related information,and a PCNN-FS was proposed to simplify the established deep learning model.After eliminating the influence of topography using CMSSA,five feature bands sensitive to mold information were selected using its feature selection module,and then a lightweight point-centered convolutional neural network(Lightweight PCNN,LPCNN)with the selected feature bands as input was constructed.The classification accuracies of the LPCNN based on PCNN-FS were 98.96% and 97.50% at pixel-wise for training and test images,respectively,and their corresponding kernel-scale classification accuracies were 98.39%and 97.12%,respectively.Compared to the full-wave PCNN model,LPCNN achieves similar classification performance while reducing about 1/3 of the model parameters and further reducing the model hardware requirements and the time required for training and testing.The overall results indicate that for moldy peanut identification spatial information is effective supplementary information,and deep spectral-spatial fusion features extracted by multilayer convolutional neural networks are more effective.The LPCNN based on PCNN-FS achieved high classification performance while reducing model complexity,which indicated that PCNN-FS could efficiently extract deep spectral-spatial fusion features and achieve fast and high accuracy identification of moldy peanuts. |
PDF Full Text Request