Research On Modeling Technology For Optical Non-destructive Detection Of Agricultural Product Quality

The advancement of optical instruments-based quality detection technology for agroproduct quality is a crucial step in constructing the agro-product safety standards,quality and safety supervision,as well as in realizing digital,networking,and smart agriculture.Optical detection technology(such as hyperspectral imaging and near-infrared analysis technologies)is essentially soft-sensing modeling,which constructs a data-driven model for connecting spectral information and quality parameters(such as defect type and distribution,physical and chemical measures)of agricultural products.However,several factors,including highdimensional characteristics,the difficulty of feature extraction caused by the multi-scale characteristics of the detected samples and the interested objects,the small-size training dataset caused by the limitation of capturing physical and chemical measures,and the potential domain shift between training and testing samples,will reduce the accuracy of the data-driven model in detecting agro-product quality parameters.This thesis explores the data-driven modeling of the external and internal quality of agro-products and studies modeling research from the perspective of computer vision and spectral quantitative analysis.The main research contents are as follows.1.To address detection problems of small-size objects caused by uneven reflected intensity on the surface of irregular agro-product,the interference of stains,and easy loss of features,a detection model combining a supervised multiple threshold segmentation model and a Canny edge detector is proposed.This method uses genetic programming(GP)algorithm to generate segmetation model.The segmentation model transforms the raw multi-spectral(MS)image into a gray image with enhanced contrast between defect and background,which eliminates the semantic aliasing between defects at different positions caused by uneven reflected intensity,and then segments the initial region of defects.Canny edge detector is used to extract the accurate defects’ edges from MS average gray image.The initial defect regions and the edge information are fused to complete the detection of small-size defects.Evaluations of the edible potato and breeding potato datasets validate the effectiveness of the proposed method.2.Defect detection of agro-products often encounters the problem of low accuracy caused by the coexistence of multi-type and multi-size defects.To tackle the detection problem of multi-type and multi-size defects in agro-products,this thesis proposes a multi-scale object detection network.The network cascades one convolutional module and six Res2 Net modules to build a backbone network for extracting multiple scale convolutional features,each of which fuses fine-grained spatial information and semantic information of defects.Two scale convolution features are inputted into the neck network and the head network separately to locate and classify multiple types of defects accurately.Experimental results on a potato MS image dataset containing five types of defects justify that the proposed network can accurately detect multi-type defects.3.To solve the insufficient accuracy and generalization ability of the spectral quantitative analysis model caused by information redundancy and multicollinearity of high-dimensional agro-product spectral data,this thesis proposed a GP-based modeling method.This method used multi-dimensional trees to encode the raw spectral data to reduce the dimensionality of spectral data while extracting various features,and then applied the multiple linear regression(MLR)method to build a prediction model.The information interaction between the multidimensional tree and the MLR model and the update of parameters in the two components are completed through individual evolution.Experimental results on six different types of spectral datasets verify that the proposed models can accurately predict the agro-product quality parameter.4.To tackle the problem that the single-task prediction model is difficult to meet the demand for accurate and fast measurement of agro-product multiple quality parameters,this thesis proposed a novel multi-task GP algorithm,to detect multiple quality parameters of agroproducts simultaneously.In this approach,the multi-dimensional trees are used to encode the raw NIR spectrum to shared features of multiple quality parameters;for each quality parameter,LS-SVR modeling is performed on the shared features to obtain private features and prediction model;during the optimization process,a new algorithm is developed to optimize the previously obtained shared and private features,and LS-SVR prediction models through population evolution by combining the M3 GP algorithm with nondominated sorting method.Experimental results on apple(two quality parameters)and sugar beet(three quality parameters)spectral datasets show that the proposed method is competitive and effective in solving the problem of multiple quality parameters predictions using the NIR spectrum.5.To tackle the problem that the performance degradation of the source model caused by domain shift,this thesis proposed a self-supervised transfer learning approach.The method firstly trained a spectral encoder consisting of three external attention modules to extract multiscale features from unlabeled source domain samples under self-supervised learning framework;secondly,it transfers the pre-trained spectra encoder followed by a feature fusion layer and an prediction head network to build a prediction model;finally,the proposed method refines the model parameters using a portion of the labeled target domain samples to adapt to unseen target domain samples.Experimental results on the apple dataset(transfer between different batches),tablet dataset(transfer between NIR instruments),and melamine dataset(transfer between product recipes)datasets validate that the proposed method has the potential to be a generic framework for tackling the common problem of domain shift-induced performance degradation of prediction models in the domain of spectrum-based quantitative analysis.