光谱学与光谱分析, 2020, 40 (9): 2826, 网络出版: 2020-11-26  

高光谱成像的图谱特征与卷积神经网络的名优大米无损鉴别

Non-Destructive Identification Method of Famous Rice Based on Image and Spectral Features of Hyperspectral Imaging With Convolutional Neural Network
作者单位
安徽大学, 国家农业生态大数据分析与应用工程研究中心, 安徽 合肥 230601
摘要
名优大米含有更多的营养价值与更高的经济价值, 不法商家为赚取更多利益, 对优质大米掺假甚至以次充好, 损害了消费者利益和大米贸易, 打击了生产者的生产积极性。 希望发展一种基于高光谱成像的图谱特征与深度学习网络的名优大米无损鉴别方法。 首先, 采集了全国具有代表性的七种名优大米400~1 000 nm范围高光谱图像, 并提取了每种大米的光谱、 纹理与形态特征。 使用多元散射校正算法做光谱预处理消除光谱散射。 连续投影算法(SPA)、 竞争自适应重加权算法(CARS)以及两者级联方法(CARS-SPA)被用来选取光谱特征的重要波长; 用 SPA选择形态、 纹理特征的重要变量。 最后, 使用深度学习网络-卷积神经网络(CNN)融合各类特征构建大米种类识别模型, 而K-近邻(KNN)、 随机森林(RF)用于与CNN模型相对比。 实验结果显示, 根据全光谱构建的模型的分类准确度达到80%以上; 其中, KNN建模效果最差; RF的效果较好; CNN网络的模型性能最优, 训练集的分类准确度(ACCT)为92.96%, 预测集的分类准确度(ACCP)为89.71%。 而重要波长光谱与全光谱相比, 分类准确度相差较多。 为进一步提升大米种类鉴别的准确度, 选用纹理、 形态两种图像特征与光谱特征进行融合, 最优结果来自光谱与形态特征重要变量所构建的模型。 其中, KNN的ACCT和ACCP分别为69%和67%; RF模型的ACCT=99.98%和ACCP=89.10%; CNN模型的效果最佳, ACCT和ACCP为97.19%和94.55%。 此外, 光谱与纹理融合的分类效果差于光谱, 说明纹理特征弱化了分类结果。 对于分类模型来说, CNN的性能明显优于两种机器学习方法, 可以提供更好的分类效果。 总而言之, CNN融合光谱与形态特征重要变量可实现对名优大米种类的准确鉴别, 这种方法有望拓展到其他农产品的分级, 种类区分和产地鉴别。
Abstract
High-quality rice contains more nutritional value and higher economic value. In order to earn more benefits, some unscrupulous merchants have adulterated high-quality rice or even replaced it with low-quality rice, which has harmed consumer interests and rice trade, and has hurt producers Production motivation. This paper hopes to develop a method for non-destructive identification of high-quality rice based on the features of images and spectra of hyperspectral imaging and deep learning networks. First, hyperspectral images in the 400~1 000 nm range of seven representative rice varieties in China were collected, and the spectra, texture, and shape features of each type of rice were extracted. The spectral features were pre-processed using the multiple scattering correction algorithms to eliminate spectral scattering. Successive projections algorithm (SPA), competitive adaptive weighting algorithm (CARS) and their cascade method (CARS-SPA) were used to select important wavelengths of spectral features. Important variables of shape and texture features were selected using SPA. Finally, convolutional neural network (CNN) was applied to fuse the above-mentioned various features to build rice varieties recognition model, while K-Nearest Neighbors (KNN) and Random Forest (RF) were used for comparison and analysis. The experimental results showed that the classification accuracy of the model constructed using the full spectroscopy reached more than 80%. Among them, KNN had the worst modeling effect and RF had a better effect. In particular, the performance of the CNN model was the best, with training set classification accuracy (ACCT) of 92.96% and prediction set classification accuracy (ACCP) of 89.71%. Compared with the full spectroscopy, the spectroscopy of the important wavelengths had worse classification accuracy. In order to further improve the accuracy of rice varieties identification, texture and shape were combined with spectral features, and the optimal result came from the model constructed of important variables of shape and spectroscopy. Among them, ACCT and ACCP of KNN were 69% and 67%, respectively. The RF model accuracy corresponded to ACCT=99.98% and ACCP=89.10%. The CNN model performed best with ACCT and ACCP of 97.19% and 94.55%. In addition, the classification effect of spectroscopy and texture fusion was worse than using only spectroscopy, indicating that texture features weakened the classification result. For classification models, the performance of CNN was significantly better than the two machine learning methods, which could provide better classification results. All in all, the important variables of shape and spectroscopy combined with CNN models could accurately identify high-quality rice varieties. The proposed method can also be applied to the identification of the variety, attribution and grade of other agricultural products.

翁士状, 唐佩佩, 张雪艳, 徐超, 郑玲, 黄林生, 赵晋陵. 高光谱成像的图谱特征与卷积神经网络的名优大米无损鉴别[J]. 光谱学与光谱分析, 2020, 40(9): 2826. WENG Shi-zhuang, TANG Pei-pei, ZHANG Xue-yan, XU Chao, ZHENG Ling, HUANG Lin-sheng, ZHAO Jin-ling. Non-Destructive Identification Method of Famous Rice Based on Image and Spectral Features of Hyperspectral Imaging With Convolutional Neural Network[J]. Spectroscopy and Spectral Analysis, 2020, 40(9): 2826.

关于本站 Cookie 的使用提示

中国光学期刊网使用基于 cookie 的技术来更好地为您提供各项服务,点击此处了解我们的隐私策略。 如您需继续使用本网站,请您授权我们使用本地 cookie 来保存部分信息。
全站搜索
您最值得信赖的光电行业旗舰网络服务平台!