基于高光谱图像和3D-CNN的苹果多品质参数无损检测

Non-destructive detection of apple multi-quality parameters based on hyperspectral imaging technology and 3D-CNN

[目的]为解决水果品质无损检测中成本、效率、精度问题,提出了一种基于高光谱图像和三维卷积神经网络(3D-CNN)的苹果高光谱多品质参数同时检测方法。[方法]使用高光谱成像系统获取400~1000 nm波段的苹果样本的高光谱反射图像并使用S-G平滑法对原始图像进行去噪处理,在此基础上,对采集到的高光谱图像通过多感兴趣位置的选取以及间隔波段抽取重组的方法进行样本扩充,再利用三维卷积神经网络建立样本扩充后的苹果高光谱图像与苹果糖度、硬度、含水量的多任务学习模型,通过该模型实现对苹果的糖度、硬度、含水量等品质参数的无损检测。[结果]采集245个苹果的高光谱图像及其对应的品质参数信息,通过样本扩充的方法将原始数据集扩充至9800个样本后进行建模和验证。结果表明:本算法建立的苹果糖度、硬度、水分的分类模型,在糖度类间隔为1°Brix、硬度类间隔为0.5 kg·cm-2、含水量类间隔为10%的情况下,糖度、硬度、水分的预测准确率分别为93.97%、92.29%和93.36%,回归模型糖度、硬度和水分的相关系数最高分别达到0.827、0.775和0.862,比最优的传统算法分别提高15.0%、17.0%和17.2%。[结论]本算法能够较准确实现苹果高光谱多品质参数同时检测,且相对传统方法预测精度有较大提升。

[Objectives]In order to solve the problem of cost,efficiency and precision in non-destructive detection of fruit quality,a simultaneous detection method of apple hyperspectral multi-quality parameters based on hyperspectral imaging technology and 3D convolutional neural networks(3D-CNN)was proposed.[Methods]First,the hyperspectral images of the apple sample in the 400-1000 nm band were acquired by the hyperspectral imaging system and the original images were denoised by the S-G smoothing method.Then,sample expansion was performed on the acquired hyperspectral images through the selection of multiple locations of interest and interval band extraction and recombination.Finally,the 3D-CNN was used to establish a multi-task learning model between expanded apple hyperspectral images and brix,firmness and moisture.Through this model,apple’s brix,firmness moisture and other quality parameters can be tested nondestructively.[Results]The hyperspectral images and the corresponding quality parameter information of 245 apples were collected.The original data set was extended to 9800 samples by sample expansion method.The results showed that the classification model of apple’s brix,firmness and moisture established by this algorithm had achieved the prediction accuracy of 93.97%,92.29%and 93.36%respectively with the brix interval of 1°Brix,the firmness interval of 0.5 kg·cm-2 and the moisture interval of 10%.At the same time,the correlation coefficient of the regression model of apple’s brix,firmness and moisture established by the algorithm respectively reached 0.827,0.775 and 0.882,which were 15.0%,17.0%and 17.2%higher than the optimal traditional algorithms.[Conclusions]The algorithm can accurately realize the simultaneous detection of multi-quality parameters of apple hyperspectral,and the prediction accuracy is higher than that of traditional methods.