基于高光谱成像的苹果果梗完整性识别方法研究

摘要

In order to achieve the grade of apple quality evaluation, it is essential to detect the integrity of the apple fruit stem. In this paper, Apple samples of 260 (stem intact fruit of 104, incomplete fruit of 92 and no fruit stem fruit of 64) were selected. The spectral information of the region of interest was extracted by hyperspectral imaging. The Stepwise multiple linear regression (SMLR) was used to extract 5 characteristic wavelengths (594.78, 572.08, 599.84, 626.48, 666nm) from the whole wavelength band (450~970nm), and the successive projections algorithm (SPA)was used to extract 7 characteristic wavelengths (467.5, 672.4, 695.4, 730.2, 840.4,952.6, 969.6nm)from the whole wavelength band (450~970nm). And then the 4 texture features of energy, entropy, moment of inertia and correlation were extracted in the region of interest. The three sets data, spectral characteristics, texture features and spectral characteristics combined with texture features which were as the input vector of support vector machine (SVM) and BP artificial neural network (BPANN),were used to identify the integrity of apple fruit stem. Five groups of cross-validation were done using cross-validation, each group separately from the training set and validation set out 30 to exchange, the average of the correctness of the five sets of validation was as the evaluation index. The experimental results showed that the fruit stalk recognition effect was not good when texture features were used as input vectors alone. SVM model accuracy rate was 50.3%, BPANN model accuracy rate was 37.2%. Fusion spectral features and texture features as input vector recognition effect in general, the correct rate was 86.6% with SPA-T-SVM model, and 80.3% with SPA-T-BPANN model, 77.2% with SMLR-T-SVM model and 74.9% with SMLR-T-BPANN model 74.9%. Only using the spectral features as the input vector, the recognition effect was better. The SPA-SVM method had the best recognition accuracy rate (91.7%), and the data calculation was small. This research provides the theoretical basis for grade evaluation of apple quality.%为了实现对苹果品质的等级评价,对果梗的完整性进行检测是必不可少的.选取了苹果样本260个,其中果梗完整果104个、果梗不完整果92个、没有果梗果64个.利用高光谱成像技术提取苹果感兴趣区域光谱信息.采用逐步多元回归算法(stepwise multiple linear regression,SMLR)从全波段(450~970nm)提取了5个特征波长(594.78,572.08,599.84,626.48,666nm),采用连续投影算法(successive projections algorithm,SPA)从全波段(450~970nm)提取了7个特征波长(467.5,672.4,695.4,730.2, 840.4,952.6,969.6nm).然后,提取感兴趣区域的惯性矩、相关性、能量和熵4个纹理特征.将光谱特征、纹理特征、光谱特征结合纹理特征3组数据,分别作为输入矢量,采用支持向量机 (support vector machine,SVM)、BP人工神经网络 (BP artificial neural network,BPANN)对苹果果梗完整性进行识别,使用交叉验证的方法进行5组交叉验证,每一组分别从训练集和验证集中拿出30个进行互换,将5组验证集正确率的平均值作为评价指标,并且对比分析结果.结果表明:单独使用纹理特征作为输入矢量,果梗识别效果不佳,SVM模型正确率为50.3%,BPANN模型正确率为37.2%.融合光谱特征和纹理特征作为输入矢量识别效果一般, SPA-T-SVM模型正确率为86.6%,SPA-T-BPANN模型正确率为80.3%,SMLR-T-SVM模型正确率为77.2%,SMLR-T-BPANN模型正确率为74.9%.只采用光谱特征作为输入矢量识别效果较好,其中SPA-SVM方法识别效果最好,识别正确率达到91.7%,且数据计算量小.该研究为苹果品质等级评价提供了理论依据.