高光谱成像的机采籽棉杂质分类检测

Classification of Impurities in Machine-Harvested Seed Cotton Using Hyperspectral Imaging

机采籽棉杂质分类检测为调整棉花清理机械加工参数和工序提供参考依据,对提升皮棉品质具有重要意义。但由于籽棉棉层分布不均匀,使得图像检测难度增大,使用传统的检测方法无法有效检测各类杂质。采用高光谱成像方法对机采籽棉中的棉叶、棉枝、地膜和铃壳(内外)五种杂质进行分类判别检测。首先采集120个机采籽棉样本的高光谱图像,选取感兴趣区域获取平均光谱曲线。发现由于物质构成的差异,不同杂质体现出不同的吸收和反射特性,不同种类物质之间的光谱差异大于同类物质。对提取的平均光谱曲线进行主成分分析(PCA),结果显示棉花、残膜和铃壳外与其他三类相比,有较好的聚集性和可分性,但是棉叶、铃壳内和棉枝三类相互叠加在一起,空间分布存在严重交叉重叠。以提取的平均光谱曲线为训练样本,选择线性判别分析(LDA)、支持向量机(SVM)和神经网络(ANN)三种分类判别算法,对算法参数进行寻优,并建立机采籽棉杂质分类判别模型。其中,经过LDA模型降维后的样本空间较PCA表现出了更好的聚集性和可分性,采用正则化防止过拟合,得到训练集准确率为86.4%,测试集准确率为86.2%;SVM模型的参数寻优结果为C=105,g=0.1,其训练集准确率为83.42%,测试集准确率为83.40%;ANN模型参数寻优得到隐含层数和神经元个数分别为2和17,训练集准确率为82.9%,测试集准确率为81.8%。对三种模型的分类效果和检测用时进行比较,LDA模型结果最优。通过对高光谱图像进行像素等级分类判别,结果显示棉花识别效果较好,植物性杂质都被有效检测,但是地膜和棉花存在误识别,分类效果与杂质光谱的分类判别模型结果一致。因此,采用高光谱成像技术可以快速、无损的检测和识别籽棉杂质,为棉花加工装备提供反馈参数,对棉花加工机械化和智能化有重要意义。

The classification and detection of impurities in machine-harvested seed cotton provides a reference for adjusting cotton cleaning mechanical processing parameters and has important significance for improving lint quality. However, the uneven distribution of seed cotton makes image detection more difficult, and traditional detection methods cannot effectively detect various impurities. The hyperspectral imaging method was used to discriminate the five impurities(cotton leaf, cotton stem, plastic film, hull inner, and hull outer) in the machine-harvested seed cotton. The hyperspectral images of 120 machine-harvested seed cotton samples were collected, and the region of interest was selected to obtain the average spectral curve. Due to the difference in the composition of materials, various impurities showed different spectral absorption and reflection characteristics, and the spectral difference of the characteristics of different materials was greater than that of similar materials. Principal component analysis(PCA) of the extracted average spectral curve showed that cotton, plastic film and hull outer were better separable than the other three types. However, the spectral distributions of cotton leaf, hull inner, and cotton stem overlapped seriously. Based on the extracted average spectral curve as the training sample, three discrimination algorithms of linear discriminant analysis(LDA), support vector machine(SVM) and neural network(ANN) were used to optimize the algorithm parameters and finally established the impurity detection model. Among them, the sample space after dimensionality reduction of the LDA model shows better separability than PCA. Regularization was used to prevent overfitting in LDA, and the accuracy rate of the training set was 86.4%, and the accuracy of the test set was 86.2%. The parameter optimization result of the SVM model was C=105, g=0.1. The accuracy of the training set was 83.42%, and the accuracy of the test set was 83.40%. The parameter optimization result of the ANN model was that the number of hidden layers and neurons were 1 and 10, respectively. The accuracy rate of the training set was 82.9%, and the accuracy rate of the test set was 81.8%. Comparing the classification accuracy and detection time of the three models, the results of the LDA model were all the best. Through the pixel level discrimination of hyperspectral images, the results show that both cotton and botanical impurities were effectively detected. However, there were misidentifications between plastic film and cotton, which was consistent with the results of the impurity spectrum classification discrimination model. Therefore, hyperspectral imaging technology can detect and identify seed cotton impurities quickly and non-destructively and provide feedback parameters for cotton processing equipment, which is of great significance to the mechanization and intelligence of cotton processing.