基于深度学习的花生高光谱图像分类方法研究

Research on Peanut Hyperspectral Image Classification Method Based on Deep Learning

利用高光谱成像技术对不同品种的花生进行快速无损分类。选取五种不同品种的花生,分别为东北小花生、富硒黑皮花生、花育36号、鲁花01号、鲁花09号,每种15颗,共75颗花生作为样本,采集400nm-1000nm波长范围内的高光谱图像,随机将6个特征波段(416nm、518nm、572nm、633nm、746nm、928nm)下的450个样本图像以2:1的比例分成训练集和测试集,建立基于深度学习的卷积神经网络模型。实验中所采用的网络模型为具有22层深度网络的GoogleNet模型,其中将dropout_ratio修改为0.6,训练集最终准确率为96%,测试集平均准确率为93.3%,每种花生的识别率均在90%及以上。最后与传统光谱处理方法PLS-DA进行对比,发现基于深度学习模型的识别率明显优于PLS-DA,结果表明,利用深度学习方法对花生快速无损分类具有可行性。

Non-destructive testing of different varieties of peanuts was performed using hyperspectral imaging techniques.Five different varieties of peanuts were selected,namely,Northeast Peanut,Selenium-enriched Black Peanut,Huayu 36,Luhua No.01 and Luhua No.09,each of which was a total of 75 peanuts as samples and collected at 400 nm.Hyperspectral imagery in the 1000 nm wavelength range,randomly divided 450 sample images in 6 characteristic bands(416 nm,518 nm,572 nm,633 nm,746 nm,928 nm)into training set and test set in a ratio of 2:1,based on depth Learning convolutional neural network model.The network model used in the experiment is a GoogleNet model with a 22-layer deep network.The dropout_ratio is modified to 0.6,the final accuracy of the training set is 96%,and the average accuracy of the test set is 93.3%.The recognition rate of each peanut is 90%and above.Finally,by comparing the traditional spectral processing method PLS-DA with the deep learning method in the SIMCA environment,it is found that the results of the deep learning method in the feature band are significantly better than the PLS-DA.