基于高光谱成像的绿皮马铃薯检测方法

针对任意放置姿态下的轻微绿皮马铃薯难以检测的问题,进行了半透射与反射高光谱成像方式的不同检测方法比较研究,最终确定较优高光谱成像方式的检测方法。分别以半透射与反射高光谱成像方式对图像维提取RGB、HSV和Lab空间颜色信息,并采用等距映射、最大方差展开、拉普拉斯特征映射进行图像信息降维;分别以半透射与反射高光谱成像方式对光谱维提取感兴趣区域的平均光谱数据,并采用局部保持投影、局部切空间排列、局部线性协调进行光谱信息降维;然后分别建立不同高光谱成像方式下的图像与光谱信息的深度信念网络模型;对识别率良好的模型采用多源信息融合技术进一步优化,并建立基于图像和光谱融合或不同成像方式融合的模型。结果表明,基于半透射和反射高光谱的光谱信息融合模型最优,校正集和测试集识别率均达到100%,可实现轻微绿皮马铃薯的无损检测。

Detection Method of Green Potato Based on Hyperspectral Imaging

To solve the problems of difficulties in detecting the slightly green potatoes placed randomly, two detection methods were compared based on the semi transmission and reflection hyperspectral imaging technologies and then a more optimal detection method was determined. 225 potatoes samples were selected, including 122 normal samples and 103 green samples. Semi transmission and reflection hyperspectral imaging technologies were used to extract the RGB, HSV and Lab color information from the image; the isometric mapping (Isomap), the maximum variance unfolding (MVU) and the Laplacian feature mapping (LE) were utilized to reduce the dimension of image information. Semi transmission and reflection hyperspectral imaging technologies were used to extract the average spectrum from the spectral region of interest; the linearity preserving projection (LPP), the local tangent space alignment (LTSA) and the locally linear coordination(LLC) were utilized to reduce the dimension of spectral information. The deep belief networks (DBN) model which is a kind of deep learning approach was developed based on the image and spectrums of different hyperspectral imaging ways. The multi source information fusion technology was used to optimize the model with a high detection accuracy and different detection models were built based on different ways of imaging or the fusion of image and spectrum. The results show that the fusion model, which is developed based on the semi transmission hyperspectral imaging and the reflection hyperspectral imaging, is the best option. Its detection rate can reach 100% in both the calibration and the validation. Non distractive detecting of the slightly green potatoes can be realized with this fusion model.