基于深度卷积神经网络的葡萄新梢图像分割  

Segmenting Grape Shoot Images Using Convolutional Neural Networks

作　　者：王书志[1] 宋广虎 冯全 WANG Shu-zhi;SONG Guang-hu;FENG Quan(College of Electrical Engineering,Northwest University for Nationalities,Lanzhou 730030,China;School of Mechanical and Electrical Engineering,Gansu Agricultural University,Lanzhou 730070,China)

机构地区：[1]西北民族大学电气工程学院,兰州730030 [2]甘肃农业大学机电工程学院,兰州730070

出　　处：《沈阳农业大学学报》2019年第4期454-462,共9页Journal of Shenyang Agricultural University

基　　金：国家自然科学基金项目(61461005);甘肃农业大学研究生重点课程建设项目(GSAU-ZDKC-1804)

摘　　要：近几年深度卷积神经网络在很多图像任务,诸如目标检测、图像分类、图像分割等方面得到了广泛应用,在图像分割方面,基于深度学习分割性能已全面超越了传统的分割算法。很多病害都会在葡萄的新梢上产生病症,在图像中准确分割出新梢,可提高病害诊断的精度。为了实现对自然条件下拍摄的葡萄新梢图像的准确分割,用相机、手机分别在不同的光照和环境条件拍摄了葡萄的新稍图像,在制作的训练图像集上对SegNet、FCN和U-NET3种卷积神经网络进行迁移学习,得到3种分割网络模型,分别用这些模型对测试集中不同环境下拍摄的新梢图像进行分割试验。在模型训练的初始阶段设置较大的学习率,以期快速到达最优解附近,随后逐步降低学习率,得到最优解。以人工分割为基准,对3种网络的分割效果进行评价。结果表明:在优选的训练模式下,3种分割网络在标准测试集T1上分割精度(MCC)达到83.58%、93.85%和89.44%,对于标准测试集T1和T2中的阴天图像,3种网络的平均MCC分别比晴天高5.42%、0.73%和0.65%。3种网络中,FCN的总体分割效果最优,在标准测试集T1上的平均分割精度(MCC)分别比SegNet和U-NET高10.27%和4.42%;从人的直观观察也可以看出,FCN分割的葡萄新梢图像轮廓光滑、视觉效果较好。光照对分割效果影响显著,阴天拍摄图像的分割效果整体好于晴天分割效果。在扩展数据集上,3种网络的分割精度均出现一定程度的下降,对于大田条件下(T3)和温室条件下(T4)手机拍摄的图像,FCN的平均分割精度(MCC)依然分别达到78.06%和74.82%,说明FCN的泛化性能较好。In recent years, deep convolutional neural networks (CNN) have been widely applied in many image tasks, such as object detection, image classification, and image segmentation. The performance of segmentation based on CNN has completely surpassed that of traditional segmentation algorithms. Many diseases infect the shoots of grape, and the accurate segmenting shoots in an image helps to improve the accuracy of disease diagnosis. To explore the best convolutional neural network architecture for accurately segmenting images of the grape shoot problem with few training data, we compared three architectures, namely, SegNet, FCN and U-NET, performed transfer learning on the training setsby fine tuning the top layers of pretrained deep networks.The training and test setswere composed of the images taken by camera or mobile phone under different weathers and environments. Three segmentation models corresponding to the above CNN were obtained after training. In early stage of the training process, the learning rate was set to a larger value, it was expected to reach the vicinity of the optimal solution quickly. Then the rate was gradually reducedto get the optimal solution. The models were used to segment the images of grape shoots in the test sets and their segmentation performances were evaluated. The resultsin the preferred training mode showed that average MCC of three networks on test image set T1 was 83.58%, 93.85% and 89.44% respectively. Lighting had a significant impact on the segmentation performance.The average MCC of three networks on the cloudy images on the standard test set A and B was 5.42%, 0.73% and 0.65% higher than that of sunny day, respectively. Among three networks, the segmentation accuracy of the FCN was best, and its MCC on the standard test set T1 was 10.69% and 1.68% higher than SegNet and U-NET, respectively. From the human visual observation, it could be seen that the shoots images segmented by FCN have a smooth contour and good visual effect. On the extended dataset, the segmentation accurac

关 键 词：自然光照 葡萄新梢 图像分割 深度学习 

分 类 号：TP1818[自动化与计算机技术—控制理论与控制工程]