基于深度信念网络的苹果霉心病病害程度无损检测

针对现有霉心病无损检测只能检测出有无病害,无法对病害程度进行判断的问题,研究并提出一种基于深度信念网络(deep belief net,DBN)的无监督检测模型。该模型由多层限制玻尔兹曼机(restricted Boltzmann machine,RBM)网络和1层反向传播(back propagation,BP)神经网络组成,RBM网络实现最优特征向量映射,输出的特征向量由BP神经网络对霉心病病害程度分类。对225个苹果样本在波长200~1 025 nm获取其透射光谱后,根据腐烂面积占横截面比例将霉心病害程度分为健康、轻度、中度和重度4种,分别用150个和75个样本作为训练集和测试集,以全光谱数据和基于连续投影算法提取的特征波长数据为输入构建病害程度判别模型,并比较DBN模型与偏最小二乘判别分析、BP神经网络和支持向量机模型的识别效果,实验结果表明,DBN模型病害判别准确率达到88.00%,具有较好的识别效果。

Non-Destructive Detection of Moldy Core in Apple Fruit Based on Deep Belief Network

Apple moldy core is a major disease affecting the internal quality of apple fruit. However, due to the lack of effective means to accurately detect moldy core in apple fruits, detection of moldy core in apples has become a major problem to be solved in the apple industry. To date, there have been no reports on the use of spectroscopy for distinguishing various degrees of moldy core decay in apple fruits. The objective of this study was to develop a non-destructive method for the detection of various degrees of moldy core decay in apple fruits using near infrared transmittance spectroscopy, successive projections algorithm (SPA), and multi-class classification algorithms partial least square-discriminant analysis (PLS-DA), back propagation artificial neural network (BP-ANN), support vector machine (SVM) and deep belief network (DBN). For developing a model to determine the degree of moldy core in apples, 225 samples were selected including a training set of 150 samples and a test set of 75 samples. The model consisted of several layers of restricted Boltzmann machine (RBM) network, which achieved eigenvector projection, and one layer of BP network, which allowed the classification of various degrees of moldy core based on the output eigenvector. The Sd value was calculated by dividing the lesion area by the total cross-sectional area. It was proposed that Sd = 0, 0 < Sd ≤ 10%, 10% < Sd < 30% and Sd ≥ 30 indicated health, mild, moderate and severe degrees, respectively. The classification accuracy of the DBN model was 88.00%, suggesting good performance of the model, and it was compared with those of the BP-ANN and SVM models.