Study On Identification Methods Of Moldy Peanut Using Hyperspectral Images

Peanut is a kind of agricultural product which is rich in yield and widely used in our country.It has high nutritional value and is an important source of vegetable oil and vegetable protein.However,peanut is prone to be contaminated with spoilage or toxigenic fungi in the process of planting,transportation and storage.Aspergillus flavus and Aspergillus parasiticus can produce secondary metabolites-aflatoxins,which is highly carcinogenic.In recent years,the frequent food contamination incidents have seriously affected the quality of food and threatened human health.Therefore,the timely and accurate detection of moldy peanuts is of great significance to ensure the food quality safety and human health in China.Traditional wet chemistry methods are time-consuming,costly,destructive to samples,and the results often deviate from the reality due to the randomness of the samples.With the development of hyperspectral technology,it can obtain the image and spectral information of the objects to be tested at the same time.It has been applied to the detection of food quality and safety by a large number of scholars,and has become an effective means for the detection of food quality and safety,plays an important role in the field of food inspection.Therefore,the theory and method of identifying moldy peanut with hyperspectral imaging technology are studied in this paper,which can provide the support of attribute and spatial position information for the mechanical selection and elimination of moldy peanut.The main contributions are as follows:1.Identification of moldy peanuts based on the spectral features of hyperspectral image.In view of the large number of bands in hyperspectral images,direct use of fullband modeling will reduce model accuracy and modeling efficiency,and is not conducive to the development of on-line monitoring system.Firstly,five common feature wavelength selection algorithms are compared,and continuous projection algorithm(SPA)is determined to be the optimal feature wavelength selection method for identifying moldy peanut using hyperspectral images.Secondly,the Ratio Spectral Index(RSI)and Normalized spectral index(NDSI)for the identification of moldy peanut were constructed based on bands of 1005.29 nm and 1056.09 nm,and then the linear discriminant analysis(LDA),quadratic discriminant analysis(QDA)and Bayes discriminant analysis(Bayes)were used to establish the discriminant models of moldy peanut.The overall classification accuracy of the model based on RSI and NDSI was similar,and the model based on LDA had the highest overall classification accuracy,which is no less than 95.24%.Finally,in order to deeply mine the spectral features sensitive to mildew information,a method based on continuous wavelet transform(CWT)is proposed to extract the spectral features of mildew information of peanut.Five continuous wavelet features(WFs)sensitive to mildew information were extracted and the recognition model of moldy peanut was constructed using support vector machine(SVM)and partial least squares discriminant analysis(PLS-DA).We find that the overall classification accuracy of the model established based on WFs is higher than that of the model established based on the key wavelengths selected by feature wavelength selection algorithms and spectral index.The CWT + SVM model has the highest precision,which shows the effectiveness of the proposed spectral feature extraction method for peanut mildew information.2.Identification of moldy peanuts based on the fusion of spatial-spectral features using hyperspectral images.Three spatial information fusion strategies are studied.The first one is preprocessing-based classification.The gray level co-occurrence matrix(GLCM)was used to extract eight texture information,which was used to establish moldy peanuts identification models using SVM and PLS-DA combined with the key wavelengths extracted by SPA.The results show that improves the classification accuracy of the model.The second on is integrated classification.Based on the extraction of SPA characteristic wavelength,JSRC are used to construct moldy peanut recognition model.The results showed that the recognition accuracy of moldy peanut was improved.The last one is postprocessing-based methods.The SVM-MRF combined with SVM and Markov Random Field(MRF)is used to construct the recognition model of moldy peanut.Comparing with three kinds of moldy peanut recognition models,the global classification accuracy of SPA+SVM-MRF model is the same as SPA+SVM model.SPA + JSRC model has the best performance constructed with spatial information,which shows that the model based on SPA and JSRC can effectively fuse local spatial and spectral information.3.Identification of moldy peanut based on deep learning.Firstly,an algorithm of identifying moldy peanut based on deep belief network(DBN)was presented and moldy peanut recognition model based on spectral information and DBN was constructed.The results show that the overall classification accuracy of the model is 97.96% or more,which is better than SPA + SVM and CWT + SVM,demonstrated the validity of the recognition model based on spectral information and DBN for moldy peanut.Secondly,a deep learning network model for moldy peanut identification is proposed based on residual multiple receptive field fusion block(ResMRFF).The overall classification accuracy of proposed deep learning model is higher than 98% on each data set,which is higher than that of SPA + SVM,SPA + JSRC and DBN.The validity of the proposed model and the great potential of depth learning in identifying moldy crop grains were fully demonstrated.4.The spatial and spectral resolution suitability of identifying moldy peanut with hyperspectral data is studied.Firstly,the suitable spectral resolution for moldy peanuts identification was optimized.The hyperspectral data is resampled and then SVM and PLS-DA models were established based on full-band and key wavelengths selected by SPA respectively on different spectral resolution.The results show that the SVM and PLSDA models have the best performance when the spectral resolution is 33.6nm.Secondly,the suitable spatial resolution of hyperspectral images for moldy peanuts identification was optimized.The spatial resolution of hyperspectral data is resampled,and SVM and PLS-DA model were established based on the full-band.The results show that SVM has the best performance when the spatial resolution is 0.8 mm,and PLS-DA has the best performance under original spatial resolution.The conclusions can be used to guide the development of on-line detection equipment for moldy peanut.