基于改进HED网络的重叠葡萄果粒轮廓特征提取  

CONTOUR FEATURE EXTRACTION OF OVERLAPPING GRAPE FRUITBASED ON IMPROVED HED NETWORK

作　　者：张舒 苗玉彬[1] Zhang Shu;Miao Yubin(School of Mechanical Engineering,Shanghai Jiao Tong University,Shanghai 200240,China)

机构地区：[1]上海交通大学机械与动力工程学院,上海200240

出　　处：《计算机应用与软件》2023年第9期157-164,共8页Computer Applications and Software

基　　金：国家自然科学基金项目(51975361)。

摘　　要：针对复杂背景以及果粒重叠造成的葡萄果粒边界难以准确检测识别的问题,提出一种基于改进HED网络的图像边缘检测算法。算法改进HED神经网络模型,使网络中每个卷积层的输出均得到充分利用;引入骰子(Dice)系数改进原有加权交叉熵损失函数。利用非极大值抑制方式对生成的边缘预测图进行边缘细化处理。实验结果表明,改进后的模型在验证集下的ODS(Optimal Dataset Scale)和OIS(Optimal Image Scale)分别达到0.801和0.817。通过比较改进HED与Canny、HED、DeepEdge三种算法对不同光照条件下的检测结果,表明改进的HED在不同条件下Dice系数高于其他算法约0.01,准确度高于其他算法约0.04。该研究结果为葡萄表型特征提取和葡萄生长规律研究提供了参考。Aimed at the problem of difficult to accurately detect and recognize the boundaries of grapes caused by the complex background and the overlapping of fruit particles,an image edge detection algorithm based on improved HED network is proposed.The algorithm improved the HED neural network model,so that the output of each convolutional layer in the network was fully utilized.The Dice coefficient was introduced to improve the original weighted cross-entropy loss function.The non-maximum value suppression method was used to perform edge refinement processing on the generated edge prediction map.Experimental results show that the ODS(optimal dataset scale)and OIS(optimal image scale)of the improved model under the validation set reach 0.801 and 0.817 respectively.By comparing the detection results of the improved HED with Canny,HED,and DeepEdge algorithms under different lighting conditions,it is shown that the improved HED has a Dice coefficient about 0.01 higher than other algorithms under different conditions,and its accuracy is about 0.04 higher than other algorithms.The research results of this paper provide a reference for the extraction of grape phenotype characteristics and the study of grape growth rules.

关 键 词：深度学习 HED网络 葡萄 果粒轮廓 特征提取 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]