Authors

Abstract

In order to realize the rapid and non-destructive detection of fresh Cerasus Humilis’ (CH) classification, and promote the deep-processing of post-harvest fresh fruit and improve market competitiveness, this study proposed a nonlinear identification method based on genetic algorithm(GA) optimized back propagation (BP) neural network of different varieties of fresh CH fruit. “Nongda-4”, “Nongda-5” and “Nongda-7” fresh CH fruit were selected as research objects to collect their visible/near-infrared spectral data dynamically. The original spectra were preprocessed by moving smoothing(MS) and standard normal variate (SNV) methods, for the characteristic wavelengths were extracted with four dimension-reducing methods, namely principal components analysis (PCA), competitive adaptive reweighed sampling (CARS), CARS-mean impact value (CARS-MIV), and random frog (RF) algorithm. Finally, the BP prediction models were established based on full-spectrum and characteristic wavelengths. At the same time, the GA optimization was used to optimize the initial weight and threshold of the BP neural network and compared with the partial least squares discrimination analysis (PLS-DA) linear model. Through comparing the MS (7)+SNV was proved to be the best preprocessing method, the CARS-MIV-GA-BP model had the best discriminant accuracy, the prediction set accuracy was 98.76%, of which the variety “Nongda-4” and “Nongda-5” recognition rate were 100%, the variety “Nongda-7” recognition rate was 96.29%. The results show that the GA can effectively optimize the initial weights and threshold randomization of the BP neural network, improve the discrimination accuracy of CH varieties, and the CARS-MIV algorithm can effectively reduce the number of input nodes of the BP neural network model, simplify the structure of BP neural network. This study provides a new theoretical basis for the detection of fresh CH fruit classification.

Abstract in Chinese

为了实现欧李鲜果分类的快速无损检测，推动采后鲜果的精深加工及提高市场竞争力。本研究提出基于遗传算法（GA）优化BP神经网络欧李鲜果品种的非线性判别方法。以产自同一地区的农大4号、农大5号和农大7号欧李果为研究对象，动态采集光谱数据。采用移动平滑法（MS）和标准正态变量（SNV）方法对原始光谱进行预处理，分别选用主成分分析（PCA）、竞争性自适应重加权算法（CARS）、竞争性自适应重加权-平均影响值算法（CARS-MIV）、随机蛙跳算法（RF）算法对光谱数据降维，将全波段和优选出的特征波长数据作为BP神经网络输入变量，采用GA优化BP神经网络的权值和阈值，建立BP、GA-BP神经网络非线性判别模型，并与偏最小二乘判别分析（PLS-DA）线性模型比较。分析得出，MS (7)+SNV为最佳预处理方法，CARS-MIV-GA-BP判别模型最佳，预测集总正确判别率为98.76%，其中 “农大4号”和“农大5号”识别率均为100%，“农大7号”识别率为96.29%。研究表明，通过GA算法能有效地优化BP神经网络初始权值和阈值随机化，可提高欧李果品种判别精度，同时CARS-MIV算法可有效减少BP神经网络模型的输入节点数，简化BP神经网络结构。该研究为欧李果在线分类检测提供了新的理论基础。