基于高光谱成像技术的青香蕉碰撞损伤检测

针对青香蕉早期轻微碰撞损伤无法用肉眼和RGB图像识别的问题,该研究利用光谱数据与图像信息,实现青香蕉早期轻微碰伤的检测和碰伤程度区分。通过高光谱成像仪获取碰撞损伤试验样品的光谱数据和图像信息,对原始光谱数据进行预处理和异常样本的剔除。通过特征波长提取,获取特征波长下的低维图像中创面区域像素点的分布数据,同时结合全像素点下的光谱反射率数据,将其作为BP神经网络模型的训练集和测试集,建立青香蕉碰撞损伤程度界定的无损检测模型。试验结果表明,利用高光谱技术可以识别肉眼不可见的轻微碰撞损伤,形成的BP神经网络检测模型的总体识别准确率为95.06%,并且可输出碰伤等级的可视化图像。研究为开发青香蕉碰伤快速无损检测系统提供理论依据。

Collision damage detection of green bananas using hyperspectral imaging

Mechanical damage is one of the most important factors to affect the banana quality and sales. The collision damage of green bananas can also be early identified by naked eye. However, manual sorting of green bananas cannot fully meet the large-scale production in recent years, due to the time-consuming, low efficiency and easy to produce the deviation. In this study, the nondestructive detection and classification of mechanical damage of green bananas were realized to utilize the spectral and image information using hyperspectral technique. Firstly, a simulation device of collision damage was built and designed, according to the structural characteristics of green bananas. The same batch of healthy green Brazilian bananas were subjected to group impact experiments with different impact energies during simulation of collision device. Secondly, the samples were placed in the hyperspectral imager within 48h after the impact to scan and collect the sample data. The sample data after black/white and lens correction were imported into the software ENVI5.3. The region of interest (ROI) of the image was selected to obtain the average spectral reflectance data and image information on the surface of the healthy green banana and the wounds. Thirdly, savitzky golay (SG) and multiplicative scattering correction (MSC) were combined to preprocess the original spectral data, and then Monte Carlo algorithm and principal component analysis (PCA) combined with Mahalanlet distance were used to remove the abnormal samples. A series of experiments were carried out to verify the classification and grouping accuracy of samples with different damage degrees. Support vector machine (SVM), least square support vector machine (LSSVM), and particle swarm optimization-least square support vector machine (PSO-LSSVM) were utilized to process the spectral reflectance data after removing abnormal samples. The successive projections algorithm (SPA) and the competitive adaptive reweighted sampling (CARS) were used to extract the characteristic wavelengths of green bananas, whereas, the interval combination optimization algorithm (ICO) was used to verify the accuracy of the extracted characteristic wavelengths. Finally, the low-dimensional image was obtained under the characteristic wavelength. The wound area and the pixel distribution data were identified using the binarization processing, Canny edge detection and image segmentation. The training and test set data of BP neural network were used to combine with the spectral reflectance data under the full pixel points derived from ENVI5.3. A nondestructive testing model was established for the classification and grouping of mechanical damage degree of green bananas. The test results show that the invisible minor collision damage was identified by hyperspectral technology. The recognition accuracy of the BP neural network detection model was 97.53%,92.59%, 93.82%, and 96.29% for the test set of healthy, mild collision, moderate collision, and severe collision samples. The accuracy of judging the overall damage degree was 95.06%. The visual image of damage level was output for the later use. This finding can provide the theoretical support to arrange the shelf life of green banana, particularly for the instrument of hyperspectral technology to identify the damage of green banana in real time.