Authors

Abstract

In order to detect apples with minor damages quickly and efficiently, which is essential for grading of apples and improving fruit quality, a method based on hyperspectral imaging and a SVM (support vector machine) model was proposed. First, to actualize this model, black-and-white correction and brightness correction based on the near-sphere geometry were applied to the apple hyperspectral image, which reduced the noise interference in the spectral image and corrected the uneven brightness distribution so that the damaged parts of the apple were easy to detect. Second, four effective wavelengths from the full-spectrum spectral data were selected via PCA (principal component analysis) and ROC (receiver operating characteristic) curve analysis. Third, the SVM model was trained using a total of 800 sets of data, which referenced the mean brightness values of intact and damaged areas in the spectral images utilizing the effective wavelengths. Additionally, 160 sets of data were employed to test the accuracy of the damage identification model. Finally, the SVM model was trained using all the samples to identify damage in 360 sets of apple images using the effective wavelengths, and the damaged areas were marked onto the apple's visible-light image. The detection accuracy for the premium, first-class and second-class apples was 90.8%, 88.3% and 87.5%, respectively, with an average detection accuracy of 88.9%. These experimental results indicated that the developed procedures were conducive to more accurate and effective detection of minor apple damage.

Abstract in Chinese

苹果轻微损伤检测对于苹果分级和提升果品整体质量至关重要，为了快速高效地检测出苹果轻微损伤，提出了一种基于高光谱图像和支持向量机的苹果轻微损伤检测方法。首先，对采集到的苹果高光谱图像使用黑白校正、基于近球体的亮度校正，减弱光谱图像中的噪声干扰、校正苹果图像亮度分布的不均匀，使苹果的损伤区域易于检测；其次，对苹果全谱段光谱数据利用主成分分析和受试者工作特征曲线分析法选择4个有效波长（488nm、529nm、632.8nm和970nm）；再次，利用有效波长下苹果光谱图像中完好和损伤区域的平均亮度值共800组数据训练支持向量机模型，并使用160组数据对模型进行损伤判定准确率检验；最后，使用全部样本训练支持向量机模型并对360组有效波长下苹果图像进行识别，将判断为损伤的区域标记到苹果的可见光图像中。特级苹果的检测准确率为90.8%、一级苹果的检测准确率为88.3%、二级苹果的检测准确率为87.5%，平均检测准确率为88.9%。实验结果表明，利用黑白校正和基于近球体的亮度校正后处理后的苹果高光谱图像，并针对由主成分分析和受试者工作特征曲线分析法选择出的4个有效波长，通过支持向量机模型进行分类，能够准确、有效的检测出苹果的轻微损伤。