针对马铃薯损伤部位随机放置会影响检测精度的问题, 提出从正对相机、 背对相机及侧对相机三个方向, 应用透射和反射高光谱成像技术采集马铃薯图像, 进行透射和反射高光谱成像的马铃薯损伤检测比较研究。 对透射和反射高光谱图像进行独立成分(IC)分析和特征提取, 利用所得特征对反射图像进行二次IC分析, 对透射和反射光谱进行变量选择, 最终分别建立基于反射图像、 反射光谱、 透射光谱的马铃薯损伤定性识别模型; 对识别准确率高的模型做进一步优化, 采用子窗口排列分析(SPA)算法对透射光谱的特征做二次选择得到3个光谱变量, 并建立任意放置的马铃薯损伤识别最优模型。 试验结果表明, 基于反射图像、 反射光谱建立的模型识别准确率较低, 其中基于反射图像的马铃薯碰伤, 侧对相机识别准确率最低为43.10%; 基于透射光谱信息建立的模型识别准确率较高, 损伤部位正对、 背对相机的识别准确率均为100%, 侧对相机为99.53%; 马铃薯损伤识别最优模型对任意放置的损伤识别准确率为97.39%。 应用透射高光谱成像技术可以检测任意放置方向下的马铃薯损伤, 该研究可为马铃薯综合品质的在线检测提供技术支持。

The randomly placed damage parts of potato will affect the detection accuracy, this paper used transmission and reflection hyperspectral imaging technology to acquire potato images of three directions(the damage part facing to the camera, back to the camera, side to the camera), and then processed the comparative study for damage detection. Independent component (IC) analysis was used to analyze the transmission and reflection hyperspectral images and to extract the features, the resulting characteristics were used for the secondary IC analysis of the reflected images and the variable selection of the transmittance and reflectance spectroscopy. Finally, the potato injury qualitative recognition model was established based on the reflection images, the reflectance spectral and the transmittance spectral; Further optimization was done for high recognition accuracy of model, and secondary variable selection was carried out for the transmission spectrum by the Sub-window Permutation Analysis(SPA) and the optimal model for damage identification of potato randomly placed was established. The results of experiments show that the accuracy of the identification model based on the reflection image and the reflection spectrum is low, wherein the potato bruise based on the reflection images falls into the lowest recognition accuracy of 43.10% when it is side to the camera; The accuracy of the model for identification based on the transmittance spectroscopy information is the highest, the recognition accuracy with the damage part facing and back to the camera is 100%, and 99.53% when it is side to the camera. The accuracy of the optimal model for identification based on the 3 kinds of transmittance spectroscopy information of randomly placed potato is 97.39%.Then the application of transmission hyperspectral imaging technology could detect potato injury in any orientation, and the research can provide technical support for the online detection of potato quality.