Research On Modelling Methods Of Apple Quality Optical Detection Based On Active Learning

The detection and grading of apple quality plays an important role in realizing the valueadded processing of apple products and enhancing market competitiveness.In recent years,near-infrared has become an important technology in the quality detection of agricultural products and fruit because of its advantages of fast acquisition of spectral data,nondestructive and limited requirements for sample preparation.Using a certain number of labeled training samples to construct a mathematical model between near infrared spectrum data and quality parameters is the key to realize accurate apple quality detection.Because the labeling of training samples requires destructive physical and chemical experiments,how to obtain a high-precision prediction model with the lowest labeling cost is a topic worth studying.Combined with active learning method,this paper studies the sample selection in the construction and updating of apple quality prediction model.The main research is as follow:1.A training sample selection strategy based on unsupervised active learning method is proposed.Firstly,the hierarchical aggregation clustering algorithm is used to divided the diversity of sample spectral data sets;Secondly,the most representative samples are selected from each cluster to be marked(physical and chemical index values)by local linear reconstruction algorithm;Finally,a training set is constructed based on the selected samples for model training.Using the near infrared spectrum data of three apple varieties in two years,a partial least squares regression prediction model for soluble solids content and firmness was constructed based on the proposed method,The experimental results show that under the same labeling cost of 200 training samples,compared with random sampling,Kennard-Stone algorithm and joint x-y distances,the root mean square error of the prediction set of the soluble solid content detection model based on the proposed unsupervised active learning is reduced by 2.0% ～ 13.2%,and the root mean square error of the prediction set of the firmness detection model is reduced by 1.2% ～ 15.7%.2.A model updating method based on uncertainty sampling strategy is proposed to improve the prediction accuracy of apple samples from different years.Firstly,some samples are selected from the original training set based on the spectral information divergence;Secondly,combined with the similarity of input feature space spectrum and the prediction deviation of output space,some samples(updating samples)are selected from the update set of the target year to be marked by using the uncertainty sampling strategy Finally,the model is iteratively updated based on selected original training samples and updating samples.Taking three varieties of apples harvested from two years as research objects,the proposed method was evaluated.The experimental results show that: at that same cost of label 150 updating sample,Compared with the model updating methods based on random sampling,Kennard-Stone algorithm and query-by-committee algorithm,the root mean square error of the prediction set is reduced by 7.6% ～ 39.9% after updating the soluble solid content detection model based on the proposed model updating method,and the root mean square error of the prediction set is reduced by 3.4% ～ 34.0%.3.A multi-model framework combining model search and active learning was proposed to improve the prediction ability of the model for different apple varieties.Firstly,the framework uses active learning method to select the most valuable samples to train a high-precision variety discrimination and classification model,and constructs a model search mechanism;Secondly,based on the active learning method,the regression model of quality detection of different varieties apples is established.Finally,for unknown apple samples,the classification results are obtained according to the variety discrimination model,and the quality detection model of specific varieties is obtained based on the model search strategy,and the quality prediction value is obtained.The experimental results showed that the multi-model framework can reduce the influence of variety differences on the accuracy of the model.Compared with the universal model,the root mean square error of the prediction set of the multi-model framework for predicting soluble solid content is reduced by 24.5% ～ 40.6%,and the root mean square error of the prediction firmness is reduced by 21.6% ～ 46.2%.