高光谱成像的猕猴桃形状特征检测

Detection of Shape Characteristics of Kiwifruit Based on Hyperspectral Imaging Technology

猕猴桃形状特征是猕猴桃在产后分级处理过程的一项重要指标,不仅影响果实外观,也决定果实等级高低的划分。传统的形状分级方法大多采用人工分级,存在耗时长、效率低、重复性差且易受人为主观影响等问题。针对传统猕猴桃形状分级存在的问题,研究利用高光谱成像建立猕猴桃正常果和畸形果的分类检测方法。以成熟期的248个金魁猕猴桃(正常果107个,畸形果141个)作为研究样本,先利用可见-近红外高光谱成像系统采集猕猴桃样本的光谱数据,再采用主成分分析法对光谱数据进行降维,得到第一主成分图像。随后提取第一主成分图像的3个特征波长(682, 809和858 nm),并对其进行融合计算,生成新的光谱图像(融合图像)。然后利用四叉树分解算法对融合图像进行分割处理,并计算掩膜图像所对应的12组形状特征参数,结合偏最小二乘线性判别分析(PLS-LDA)、反向传播神经网络(BPNN)、最小二乘支持向量机(LSSVM)建立判别模型,对比分析,最终得到猕猴桃形状特征的最佳分类模型。结果表明,所建立的三种分类模型中, BPNN和LSSVM模型的分类效果较好,总体分类准确率均在95%以上;PLS-LDA的效果略差,训练集和测试集的总体准确率分别为80.12%和76.83%。其中BPNN模型训练集和测试集的总体分类准确率分别为98.19%和97.56%,总体误判个数分别为3和2,而LSSVM模型的总体准确率分别为97.59%和95.12%,总体误判个数分别为4和4。对猕猴桃正常果的检测,三种模型的分类效果分别为:LSSVM最好、 BPNN其次、 PLS-LDA最差。对猕猴桃畸形果的检测,三种模型的分类效果分别为:BPNN最优、 LSSVM其次, PLS-LDA效果最差。因此,猕猴桃形状特征的最佳分类模型是BPNN模型。试验结果说明,可利用高光谱成像对猕猴桃形状特征进行分类判别。为猕猴桃形状特征的快速、准确无损检测研究提供了理论支持。

The shape characteristic of kiwifruit, an important indicator in the post-harvest grading process, not only affects the appearance quality of fruits but also determines the level division of them. Most of the traditional shape grading methods were adopted manual grading, which had the disadvantages of long time-consuming, low efficiency, poor repeatability and strong subjective influence. This paper used visible and near-infrared(VIS/NIR) hyperspectral imaging technique to discriminate normal and malformed kiwifruit. Firstly, 248 mature "Jinkui" kiwifruit(107 normal samples and 141 malformed samples) were prepared. The visible-near-infrared hyperspectral imaging acquisition system(400~1 000 nm) was constructed to acquire the hyperspectral image of kiwifruit. After completing the spectral image acquisition, used principal component analysis(PCA) method to reduce dimensions and obtain the first principal component image for extracting three characteristic wavelengths(682, 809 and 858 nm). Then, the wavelengths were calculated to generate a new spectral image(fused image). Furthermore, the image was segmented by the quadtree decomposition algorithm, and the corresponding 12 sets of shape characteristic parameters were calculated based on the extracted mask images. The classification models by partial least squares-linear discriminant analysis(PLS-LDA), backpropagation neural network(BPNN), and least squares support vector machine(LSSVM) were established. Finally, compared and analyzed, the best model of kiwifruit shape characteristics was obtained. The results showed that among three classification models, BPNN and LSSVM models had better classification consequences: the overall classification accuracy was above 95%;The effects of PLS-LDA model was slightly worst: the overall accuracy of the training and test sets were 80.12% and 76.83%, respectively. Among them, the overall classification accuracy of BPNN was 98.19% and 97.56% in training and test set, respectively, and the total number of misjudgments were 3 and 2, respectively. Yet, the overall accuracy of LSSVM model was 97.59% and 95.12%, respectively, the total number of misjudgments were 4 and 4, respectively. For the classification effects of kiwifruit normal, the performances of three models were: LSSVM best, BPNN followed, and PLS-LDA bottom. For the classification effects of malformation, the performances of three models were: BPNN optimal, LSSVM followed, and PLS-LDA foot. Therefore, the best classification model for kiwifruit shape characteristics was BPNN. The experimental results showed that the shape characteristics of kiwifruit could be classified and identified and had an ideal effect. In the future, it is feasible to detect fruit shape combining the visible-near-infrared hyperspectral imaging technique. The result can provide the theoretical support for the rapid and accurate non-destructive detection of kiwifruit shape features using spectral information.