基于高光谱成像和深度学习的山核桃内源性异物检测

Inspection of Endogenous Foreign Body in Chinese Hickorynut Based on Hyperspectral Imaging and Deep Learning

山核桃壳是山核桃加工生产中的内源性异物,其颜色与果仁差异性较小,难以通过颜色进行准确识别。针对此问题,提出了一种基于高光谱成像和深度学习的山核桃内源性异物检测方法。以山核桃为研究对象,根据山核桃的组成和结构特征,将山核桃分为内仁、外仁、内壳和外壳4种组分,使用高光谱成像技术获取了各组分的高光谱图像,依次通过大津法、形态学算法和逻辑与运算对高光谱图像进行了背景分割,提取了山核桃各组分像素点的光谱,并利用多元散射校正对各组分光谱进行了预处理。基于一维神经网络(1DCNN),提取各组分光谱的深度特征,建立山核桃内源性异物的1DCNN检测模型。为了提高检测模型的性能,将归一化的各组分光谱转化为二维向量,作为二维卷积神经网络(2DCNN)的输入,建立2DCNN山核桃内源性异物的检测模型,模型的性能优于所建立的1DCNN模型,将训练集和测试集的分类正确率分别提高到100%和98.5%。

Chinese hickory shell is the endogenous foreign body in its kernel production.Since the shell and kernel is similar in color,it is difficult to distinguish the shell and kernel accurately by color.In order to solve this problem,a nondestructive method based on hyperspectral imaging and deep learning for detecting endogenous foreign body in Chinese hickory nut was proposed.According to composition and structure of hickory nut,all samples can be divided into the inner shell,outer shell,inner kernel and outer kernel groups.After the hyperspectral images of each group was collected,the background of each hyperspectral image was removed by the Otsu method morphological algorithm and logical‘and’operation.The spectra of pixels in each group were extracted and preprocessed by multiplicative scatter correction(MSC)method.The deep features of the spectra were extracted by one dimension convolutional neural networks(1DCNN)and an 1 DCNN model was established for detection of endogenous foreign body in hickory nut.To improve the detection accuracy,the spectra of pixels were normalized and reshaped into two-dimensional vector as the input of two dimension convolutional neural networks(2DCNN).The performance of the proposed 2 DCNN model was better than that of the 1DCNN model.The accuracies of 100%and 98.5%were achieved for the training set and testing set,respectively.