Abstract: In order to realize the rapid non⁃destructive measurement of jujube volume and improve the precision of jujube volume classification,a method based on depth image and BP neural network was proposed.By collecting the depth image of jujube,plane segmentation and jujube clustering were carried out by segmentation algorithm,and cylinder fitting of the jujube point cloud after clustering was carried out to establish a 3D model of jujube.Five features,including long diameter,short diameter,contour circumference,projection area and sphericity,were extracted from the 3D model of jujube by using the bounding box algorithm and convex hull method.Seven sets of BP neural network models with different feature combinations were established to predict the volume of jujube.By analyzing the prediction results of seven groups of models,the results showed that the model containing all the characteristics of jujube had the best prediction results.The determination coefficient(R²)of forecasted values and measured values of volume was 0.861 87,the root mean square error(RMSE)was 1.66 mL,and the average relative error between the predicted volume and the measured volume was 6.65%.The results demonstrate that the estimation of jujube volume by depth image and BP neural network has high prediction accuracy.

Key words: Jujube, Depth image, BP neural network, Feature extraction, Volume measurement

摘要： 为了实现红枣体积的快速无损测量，提高红枣体积分级精度，提出了一种基于深度图像和BP（Back propagation，反向传播）神经网络的红枣体积测量方法。通过采集红枣深度图像，利用分割算法进行平面分割和红枣聚类，对聚类后的红枣点云分别进行柱面拟合，建立红枣3D模型。采用包围盒算法、凸包法等多种方法提取红枣3D模型上的长径、短径、轮廓周长、投影面积、球度共5种特征，建立7组不同特征组合的BP神经网络模型，分别预测红枣体积。结果表明，包含红枣5种特征的模型预测结果最好，其体积预测值与实测值的决定系数（R²）为0.861 87，均方根误差（RMSE）为1.66 mL，与实测值的平均相对误差为6.65%。表明采用深度图像和BP神经网络估测红枣体积具有较高预测精度。

关键词: 红枣, 深度图像, BP神经网络, 特征提取, 体积测量