基于深度卷积神经网络的葡萄新梢图像分割
|
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 was5.42%, 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 accuracy of these networks all decreased to some extent. For the images taken by mobile phone in the field conditions(T3) and greenhouse conditions(T4), the average MCC of the FCN still reached78.06% and 74.82%. These results indicated that FCN could possess a good generalization performance.
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 was5.42%, 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 accuracy of these networks all decreased to some extent. For the images taken by mobile phone in the field conditions(T3) and greenhouse conditions(T4), the average MCC of the FCN still reached78.06% and 74.82%. These results indicated that FCN could possess a good generalization performance.
引文
[1]陈鸿翔.基于卷积神经网络的图像语义分割[D].杭州:浙江大学,2016.
[2]魏伟波,芮筱亭.图像边缘检测方法研究[J].计算机工程与应用,2006(30):88-91.
[3]林正春,王知衍,张艳青.最优进化图像阈值分割算法[J].计算机辅助设计与图形学报,2010,22(7):1201-1206.
[4]李俊,杨新,施鹏飞.基于Mumford-Shah模型的快速水平集图像分割方法[J].计算机学报,2002,25(11):1175-1183.
[5]夏雪.基于计算机视觉的苹果树果实探测与定位方法[D].北京:中国农业科学院,2018.
[6]刘立群,王联国,火久元,等.基于改进和声搜索算法的玉米叶片病害图像分割算法[J].计算机与应用与软件2016,33(4):183-186.
[7]张静,王双喜.温室植物病害图像处理技术中图像分割方法的研究[J].内蒙古农业大学报2017,28(3):19-22.
[8]尹丽华,李范鸣,刘士建.基于区域分割与融合的全景稳像算法[J].红外与激光工程,2018,47(9):391-400.
[9]张帅.基于深度学习的植物叶片识别算法探究[D]北京:北京林业大学,2016.
[10]张正昊.基于图像处理的牧草形状特征提取与识别系统设计[D].呼和浩特:内蒙古农业大学,2014.
[11]胡静,陈志泊,杨猛,等.基于全卷积神经网络的植物叶片分割算法[J].北京林业大学学报,2018,40(11):131-136.
[12]齐龙.基于图像处理的作物病害诊断及叶片形态参数测量技术的研究[D].长春:吉林大学,2006.
[13]李娇娇.玉米叶部病斑图像智能处理算法的研究与实现[D].北京:北京邮电大学,2012.
[14]王昱潭.基于机器视觉的灵武长枣定位与成熟度判别方法研究[D].北京:北京林业大学,2014.
[15]纪超.温室果蔬采摘机器人视觉信息获取方法及样机系统研究[D].武汉:华中农业大学,2011.
[16]李凯,张建华,冯全,等.复杂背景与天气条件下的棉花叶片图像分割方法[J].中国农业大学学报,2018,23(2):88-98l.
[17]赵金阳,冯全,王书志,等.一种改进的葡萄叶片自动分割算法[J].计算机辅助设计与图形学学报,2017,22(11):140-147.
[18] ZHOU T.AnImage recognition model based on improved convolutional neural network[J].Journal of Computational&Theoretical Nanoscience,2016,13(7):4223-4229.
[19]周云成,许童羽,邓寒冰,等.基于双卷积链Fast R-CNN的番茄关键器官识别方法[J].沈阳农业大学学报,2018,49(1):65-74.
[20] SHI J,MALIKJ.Normalizedcuts and image segmentation[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2000,22(8):888-905.
[21]顾文璇.基于深度卷积网络的农业病虫害图像分类识别应用[D].武汉:武汉轻工业大学,2017.
[22]赵兵,冯全.基于全卷积网络的葡萄病害叶片分割[J].南京农业大学学报,2018,41(4):752-759.
[23] BADRINARAYANAN V,KENDALL A,CIPOLLAR.SegNet:Adeep convolutional encoderdecoder architecture for scene segmentation[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2017,39(12):2841-2495.
[24] LONG J,SHELHAMER E,LOOG J,et al.Fully convolutional networks for semantic segmentation[J].IEEE Transactions on Pattern Analysis&Machine Intelligence,2014,39(4):640-651.
[25] SEVASTOPOISKYA.Optic disc and cup segmentation methods for glaucoma detection with modification of U-Net convolutional neural network[J].Pattern Recognition and Image Analysis,2017,27(3):618-624.
[26]李凯,冯全,张建华.棉花苗叶片复杂背景图像的联合分割算法[J].计算机辅助设计与图形学学报,2017,29(10):1871-1880.
[2]魏伟波,芮筱亭.图像边缘检测方法研究[J].计算机工程与应用,2006(30):88-91.
[3]林正春,王知衍,张艳青.最优进化图像阈值分割算法[J].计算机辅助设计与图形学报,2010,22(7):1201-1206.
[4]李俊,杨新,施鹏飞.基于Mumford-Shah模型的快速水平集图像分割方法[J].计算机学报,2002,25(11):1175-1183.
[5]夏雪.基于计算机视觉的苹果树果实探测与定位方法[D].北京:中国农业科学院,2018.
[6]刘立群,王联国,火久元,等.基于改进和声搜索算法的玉米叶片病害图像分割算法[J].计算机与应用与软件2016,33(4):183-186.
[7]张静,王双喜.温室植物病害图像处理技术中图像分割方法的研究[J].内蒙古农业大学报2017,28(3):19-22.
[8]尹丽华,李范鸣,刘士建.基于区域分割与融合的全景稳像算法[J].红外与激光工程,2018,47(9):391-400.
[9]张帅.基于深度学习的植物叶片识别算法探究[D]北京:北京林业大学,2016.
[10]张正昊.基于图像处理的牧草形状特征提取与识别系统设计[D].呼和浩特:内蒙古农业大学,2014.
[11]胡静,陈志泊,杨猛,等.基于全卷积神经网络的植物叶片分割算法[J].北京林业大学学报,2018,40(11):131-136.
[12]齐龙.基于图像处理的作物病害诊断及叶片形态参数测量技术的研究[D].长春:吉林大学,2006.
[13]李娇娇.玉米叶部病斑图像智能处理算法的研究与实现[D].北京:北京邮电大学,2012.
[14]王昱潭.基于机器视觉的灵武长枣定位与成熟度判别方法研究[D].北京:北京林业大学,2014.
[15]纪超.温室果蔬采摘机器人视觉信息获取方法及样机系统研究[D].武汉:华中农业大学,2011.
[16]李凯,张建华,冯全,等.复杂背景与天气条件下的棉花叶片图像分割方法[J].中国农业大学学报,2018,23(2):88-98l.
[17]赵金阳,冯全,王书志,等.一种改进的葡萄叶片自动分割算法[J].计算机辅助设计与图形学学报,2017,22(11):140-147.
[18] ZHOU T.AnImage recognition model based on improved convolutional neural network[J].Journal of Computational&Theoretical Nanoscience,2016,13(7):4223-4229.
[19]周云成,许童羽,邓寒冰,等.基于双卷积链Fast R-CNN的番茄关键器官识别方法[J].沈阳农业大学学报,2018,49(1):65-74.
[20] SHI J,MALIKJ.Normalizedcuts and image segmentation[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2000,22(8):888-905.
[21]顾文璇.基于深度卷积网络的农业病虫害图像分类识别应用[D].武汉:武汉轻工业大学,2017.
[22]赵兵,冯全.基于全卷积网络的葡萄病害叶片分割[J].南京农业大学学报,2018,41(4):752-759.
[23] BADRINARAYANAN V,KENDALL A,CIPOLLAR.SegNet:Adeep convolutional encoderdecoder architecture for scene segmentation[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2017,39(12):2841-2495.
[24] LONG J,SHELHAMER E,LOOG J,et al.Fully convolutional networks for semantic segmentation[J].IEEE Transactions on Pattern Analysis&Machine Intelligence,2014,39(4):640-651.
[25] SEVASTOPOISKYA.Optic disc and cup segmentation methods for glaucoma detection with modification of U-Net convolutional neural network[J].Pattern Recognition and Image Analysis,2017,27(3):618-624.
[26]李凯,冯全,张建华.棉花苗叶片复杂背景图像的联合分割算法[J].计算机辅助设计与图形学学报,2017,29(10):1871-1880.
© 2004-2018 中国地质图书馆版权所有 京ICP备05064691号 京公网安备11010802017129号 地址：北京市海淀区学院路29号 邮编：100083 电话：办公室：(+86 10)66554848；文献借阅、咨询服务、科技查新：66554700 |