【摘要】 苹果早期轻微瘀伤是影响苹果品质的重要因素。早期轻微瘀伤在可见光下肉眼无法识别，为寻找一种高效的苹果早期轻微瘀伤识别方法，以红富士苹果为研究对象，通过倒立摆装置人为制造不同程度的苹果瘀伤。首先采用傅里叶变换近红外光谱仪采集80个无损样品、 60个轻度损伤样品以及60个重度损伤样品各自在损伤后0、 10、 20和30 min的近红外漫反射光谱；采用SNV作为光谱数据预处理方法，谱区范围选择4 000～9 000 cm^-1;主成分个数为9,采用偏最小二乘法-判别分析（PLS-DA）和支持向量机（SVM）两种方法建立“无损-有损”二分类模型，预测集的平均识别率分别为85.00%和89.80%,模型识别效果有待提高。在以上实验结果的基础上，使用波段范围为1 000～2 350 nm的近红外相机采集无损、轻度瘀伤、中度瘀伤和重度瘀伤的100个苹果样品的近红外图像，相机加装1 150和1 400 nm的滤光片后分别再次采集这些苹果样品的近红外图像，所有图像均在瘀伤产生后立即采集。共采集3个波段、 4种瘀伤程度的苹果样本图像共1 200张。提取图像吸光度信息并分别建立KNN、 SVM和DT分类模型，DT法“无损-损伤”二分类模型和“无损-轻度-重度”三分类模型的识别率最高，分别为99.00%和94.67%。相比基于近红外光谱的苹果早期外部瘀伤识别方法，近红外相机多波段成像法在苹果表面早期瘀伤和瘀伤程度分类的应用中都有更高的识别准确率。与此同时近红外相机成像方法方便确定瘀伤的位置，这为苹果表面瘀伤的实时在线检测与分类提供了一种快速高效的新思路。更多还原

【Abstract】 Early light bruising in apples is an important factor affecting apple quality. Since early minor bruises cannot be identified by the naked eye in visible light, and in order to control the time of bruise production, a red Fuji apple was selected as the research object, and different degrees of apple bruises were artificially created through an inverted pendulum device. To find an efficient method to identify early minor apple bruises, Fourier transforms near-infrared spectrometer was first used to collect near-infrared diffuse reflectance spectra of 80 undamaged samples, 60 lightly damaged samples, and 60 heavily damaged samples at 0, 10, 20 and 30 min post-damage, respectively. SNV is used as the spectral data preprocessing method. The spectral range is 4 000～9 000 cm^-1, and the number of principal components is 9. The “Nondestructive-damage” classification model is established by partial least squares-discriminant analysis（PLS-DA） and support vector machine（SVM） respectively. The average recognition rate of the prediction set was 85.00% and 89.80%, respectively, and the model recognition effect needs to be improved. Based on the above experimental results, the NIR images of apples with no damage, light bruise, moderate bruise, and severe bruise of 100 apples were acquired by using NIR cameras with a wavelength range of 1 000 to 2 350 nm. The NIR images of these apples were acquired again with the addition of 1 150 and 1 400 nm filters, respectively. 1 200 images of apples in 3 bands and 4 bruise levels were acquired. The image absorbance information was extracted, and KNN, SVM, and DT classification models were built respectively. The highest recognition rates of 99.00% and 94.67% were achieved by the “nondestructive-damage” classification model and the “nondestructive-mild damage-severe damage” classification model using the DT method, respectively. Compared with the NIR spectroscopy method, the NIR camera multi-band imaging method has higher recognition accuracy in applying both early bruises and bruise degree classification on apple surfaces. At the same time, the NIR camera imaging method is convenient for determining the location of the bruise, which provides a fast and efficient new idea for real-time online detection and classification of bruises on apple surfaces.更多还原