高级检索

留言板

尊敬的读者、作者、审稿人, 关于本刊的投稿、审稿、编辑和出版的任何问题, 您可以本页添加留言。我们将尽快给您答复。谢谢您的支持!

姓名
邮箱
手机号码
标题
留言内容
验证码

基于多特征融合和CNN模型的树种图像识别研究

刘嘉政 王雪峰 王甜

刘嘉政, 王雪峰, 王甜. 基于多特征融合和CNN模型的树种图像识别研究[J]. 北京林业大学学报, 2019, 41(11): 76-86. doi: 10.13332/j.1000-1522.20180366
引用本文: 刘嘉政, 王雪峰, 王甜. 基于多特征融合和CNN模型的树种图像识别研究[J]. 北京林业大学学报, 2019, 41(11): 76-86. doi: 10.13332/j.1000-1522.20180366
Liu Jiazheng, Wang Xuefeng, Wang Tian. Image recognition of tree species based on multi feature fusion and CNN model[J]. Journal of Beijing Forestry University, 2019, 41(11): 76-86. doi: 10.13332/j.1000-1522.20180366
Citation: Liu Jiazheng, Wang Xuefeng, Wang Tian. Image recognition of tree species based on multi feature fusion and CNN model[J]. Journal of Beijing Forestry University, 2019, 41(11): 76-86. doi: 10.13332/j.1000-1522.20180366

基于多特征融合和CNN模型的树种图像识别研究

doi: 10.13332/j.1000-1522.20180366
基金项目: 国家重点研发计划项目(2017YFC0504106)
详细信息
    作者简介:

    刘嘉政。主要研究方向:林业信息技术应用。Email:liujiazheng0919@163.com  地址:100091北京市海淀区香山路东小府1号中国林业科学研究院资源信息研究所

    责任作者:

    王雪峰,博士,研究员。主要研究方向:森林资源监测与计算机视觉。Email:xuefeng@ifrit.ac.cn 地址:同上

Image recognition of tree species based on multi feature fusion and CNN model

  • 摘要: 目的在树种图像识别时会存在类内差异、类间相似的现象,因此导致基于单一人工特征的传统识别方法难以达到理想的识别效果。针对这一问题,本文基于卷积神经网络,提出一种将图像深层特征和人工特征融合的树种图像深度学习识别方法。方法将6类常见树种(樟子松、山杨、白桦、落叶松、雪松和白皮松)图像作为研究对象。首先,通过裁剪、水平翻转、旋转等操作,对原始树种图像集进行数量扩增,并划分为训练集和测试集,建立本次树种识别实验的图像库;其次,将本文模型设计为3路并列网络,分别选取RGB图像、HSV图像、LBP-HOG图像,从图像像素、色彩、纹理和形状的角度出发,对上述树种图像进行识别。一方面构建适合本文实验的CNN深度学习模型,将训练集样本中RGB图像和相对应的HSV图像作为第1路和第2路CNN模型的输入,进行树种图像深层特征提取;另一方面,对训练集进行高斯滤波去噪和人工提取LBP-HOG特征来代表纹理、形状特征,作为第3路CNN模型的输入。然后,将3路模型各自得到的特征在最后一层全连接层进行汇总,作为softmax分类器的最终分类依据。最后,为检验本文方法的可行性,利用上述特征和训练集对SVM分类器、BP神经网络以及现有的深度学习LeNet-5模型、VGG-16模型进行训练,对测试集进行识别验证,来比较最终的识别效果。结果本文提出的多特征融合CNN模型,训练准确率为96.13%,平均验证识别准确率为91.70%。基于单路训练的CNN树种识别模型中,RGB图像作为训练输入值时,识别率最高,为75.21%,HSV特征识别率次之,LBP-HOG特征最差;多特征融合情况下,基于RGB + H通道 + LBP条件下,验证识别准确率最高,达到93.50%;RGB + HSV + LBP + HOG组合识别率不增反降,识别率为89.50%。同样的特征或特征组合条件下,SVM、BP神经网络、LeNet-5模型和VGG-16模型所获得的识别率均低于本文模型的识别率。结论基于RGB + H通道 + LBP特征融合条件下,运用3路并列CNN模型,对本文6类树种图像进行识别的识别率最高,克服了在单一特征情况下识别率低的问题,识别效果也非常理想,实现了从大量不同树种图像中自动识别出具体类别。

     

  • 图  1  6类树种图像

    Figure  1.  6 tree species images

    图  2  树种图像样本数

    Figure  2.  Sample number of tree species images

    图  3  并列CNN网络结构

    Figure  3.  Parallel CNN network structure

    图  4  准确率与损失率

    Figure  4.  Training accuracy rate and loss rate

    图  5  RGB图像训练特征图

    Figure  5.  Training feature map of RGB image

    图  8  HOG图像训练特征图

    Figure  8.  Training feature map of HOG images

    图  6  HSV图像训练特征图

    Figure  6.  Training feature map of HSV images

    图  7  LBP图像训练特征图

    Figure  7.  Training feature map of LBP images

    图  9  树种识别结果的混淆矩阵

    bh:白桦Betula platyphylla;zzs:樟子松Pinus sylvestris var. Mongolica;lys:落叶松Larix gmelinii;sy:山杨Populus davidiana;bps:白皮松Pinus bungeana;xs:雪松Cedrus deodara

    Figure  9.  Confusion matrix of tree species recognition results

    表  1  本文方法的实验结果

    Table  1.   Experimental results in this paper %

    项目 Item樟子松
    Pinus sylvestris
    var. mongolica
    山杨
    Populus
    davidiana
    白桦
    Betula
    platyphylla
    落叶松
    Larix
    gmelinii
    雪松
    Cedrus
    deodara
    白皮松
    Pinus
    bungeana
    验证识别准确率
    Accuracy rate of verification and recognition
    91.50 90.40 92.80 88.70 90.10 93.50
    平均验证识别准确率
    Average accuracy rate of verification and recognition
    91.17
    下载: 导出CSV

    表  2  不同卷积核数目的训练准确率

    Table  2.   Training accuracy rate of different convolution kernel numbers

    卷积核数目
    Convolution kernel number
    训练准确率
    Training accuracy rate/%
    32-64-64-12875.26
    32-64-128-6472.21
    32-64-128-19278.28
    48-64-128-12875.12
    48-128-192-12881.79
    48-128-128-19282.23
    64-128-128-12896.13
    64-128-192-19291.27
    64-64-128-19292.02
    下载: 导出CSV

    表  3  不同特征组合的识别率

    Table  3.   Recognition rate of different feature combinations

    特征
    Feature
    训练准确率
    Training accuracy rate/%
    验证准确集
    Verification accuracy set
    验证集
    Validation set
    验证识别准确率
    Verifying the recognition accuracy rate/%
    RGB 75.21 433 600 72.17
    HSV 71.56 416 600 69.33
    HOG-LBP 56.28 314 600 52.33
    H 63.78 377 600 62.83
    S 68.12 391 600 65.17
    V 64.14 375 600 62.50
    RGB + H 78.26 451 600 75.17
    RGB + S 75.26 440 600 73.33
    RGB + V 78.21 445 600 74.17
    RGB + LBP 75.23 446 600 74.33
    RGB + HOG 77.58 460 600 76.67
    RGB + HOG + H 86.29 498 600 83.00
    RGB + HOG + S 82.15 475 600 79.17
    RGB + HOG + V 84.26 493 600 82.17
    RGB + LBP + H 96.13 561 600 93.50
    RGB + LBP + S 88.14 525 600 87.50
    RGB + LBP + V 92.36 541 600 90.17
    RGB + HSV + LBP-HOG 990.56 537 600 89.50
    下载: 导出CSV

    表  4  与其他方法的识别率比较结果

    Table  4.   Comparison results of recognition rates with other methods %

    方法
    Method
    识别率 Recognition rate
    白桦
    Betula platyphylla
    樟子松
    Pinus sylvestris var. mongolica
    落叶松
    Larix gmelinii
    雪松
    Cedrus deodara
    山杨
    Populus davidiana
    白皮松
    Pinus bungeana
    SVM 47.25 48.41 48.02 44.95 43.66 45.28
    BP 36.25 39.41 40.25 39.65 39.12 38.25
    LeNet-5 59.28 55.36 57.25 54.78 55.45 52.36
    VGG-16 63.21 60.17 66.28 64.58 64.11 60.29
    本文方法
    Method in this study
    92.80 91.50 88.70 90.10 90.40 93.50
    下载: 导出CSV
  • [1] 王昌腾. 基于应用型人才培养提高学生树木识别教改探索[J]. 现代园艺, 2018(13):165−166.

    Wang C T. Exploration on the teaching reform of improving students'tree recognition based on the cultivation of applied talents[J]. Modern Horticulture, 2018(13): 165−166.
    [2] Richter R, Reu B, Wirth C, et al. The use of airborne hyperspectral data for tree species classification in a species-rich Central European forest area[J]. International Journal of Applied Earth Observation and Geoinformation, 2016, 52: 464−474. doi: 10.1016/j.jag.2016.07.018
    [3] Pham L T H, Brabyn L, Ashraf S. Combining QuickBird, LiDAR, and GIS topography indices to identify a single native tree species in a complex landscape using an object-based classification approach[J]. International Journal of Applied Earth Observation and Geoinformation, 2016, 50: 187−197. doi: 10.1016/j.jag.2016.03.015
    [4] 陈明健, 陈志泊, 杨猛, 等. 叶片传统特征和距离矩阵与角点矩阵相结合的树种识别算法[J]. 北京林业大学学报, 2017, 39(2):108−116.

    Chen M J, Chen Z B, Yang M, et al. A tree species identification algorithm combining traditional leaf characteristics and distance matrix with corner matrix[J]. Journal of Beijing Forestry University, 2017, 39(2): 108−116.
    [5] 李可心, 戚大伟, 牟洪波, 等. 基于灰度共生矩阵与SOM神经网络的树皮纹理特征识别[J]. 森林工程, 2017, 33(3):24−27. doi: 10.3969/j.issn.1006-8023.2017.03.006

    Li K X, Qi D W, Mou H B, et al. Bark texture recognition based on gray level co-occurrence matrix and SOM neural network[J]. Forest Engineering, 2017, 33(3): 24−27. doi: 10.3969/j.issn.1006-8023.2017.03.006
    [6] 杨洋. 基于小波变换及SVM理论的树木种类识别研究[D]. 哈尔滨: 东北林业大学, 2017.

    Yang Y. Research on tree species recognition based on wavelet transform and SVM theory[D]. Harbin: Northeast Forestry University, 2017.
    [7] 于海鹏, 刘一星, 刘镇波. 基于图像纹理特征的木材树种识别[J]. 林业科学, 2007,43(4):77−81,146−147. doi: 10.3321/j.issn:1001-7488.2007.04.013

    Yu H P, Liu Y X, Liu Z B. Wood species identification based on image texture features[J]. Forestry Science, 2007,43(4): 77−81,146−147. doi: 10.3321/j.issn:1001-7488.2007.04.013
    [8] 孙伶君, 汪杭军, 祁亨年. 基于分块LBP的树种识别研究[J]. 北京林业大学学报, 2011, 33(4):107−112.

    Sun L J, Wang H J, Qi H N. Study on tree species identification based on block LBP[J]. Journal of Beijing Forestry University, 2011, 33(4): 107−112.
    [9] Bertrand S, Ameur R B, Cerutti G, et al. Bark and leaf fusion systems to improve automatic tree species recognition[J]. Ecological Informatics, 2018, 46: 57−73. doi: 10.1016/j.ecoinf.2018.05.007
    [10] Zhao Z Q, Ma L H, Cheung Y, et al. ApLeaf: an efficient android-based plant leaf identification system[J]. Neurocomputing, 2015, 151: 1112−1119. doi: 10.1016/j.neucom.2014.02.077
    [11] 赵鹏超, 戚大伟. 基于卷积神经网络和树叶纹理的树种识别研究[J]. 森林工程, 2018, 34(1):56−59. doi: 10.3969/j.issn.1006-8023.2018.01.013

    Zhao P C, Qi D W. Study on tree species identification based on convolution neural network and leaf texture[J]. Forest Engineering, 2018, 34(1): 56−59. doi: 10.3969/j.issn.1006-8023.2018.01.013
    [12] Li Q, You X, Li K, et al. Deep hierarchical feature extraction algorithm[J]. Pattern Recognition and Artificial Intelligence, 2017, 30(2): 127−136.
    [13] Lecun Y, Bottou L, Bengio Y, et al. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 1998, 86(11): 2278−2324. doi: 10.1109/5.726791
    [14] Krizhevsky A, Sutskever I, Hinton G E. ImageNet classification with deep convolutional neural networks[C]//Proceedings of Advances in Neural Information Processing Systems. Cambridge: Massachusetts Institute of Technology Press, 2012: 1106−1114.
    [15] Deng J, Dong W, Socher R, et al. ImageNet: a large-scale hierarchical image database[C]//Proceedings of the 2009 IEEE Conference on Computer Vision and Pattern Recognition. Washington D C: IEEE Computer Society, 2009: 248−255.
    [16] Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[C/OL]. arXiv, 2014[2018−05−06]. https://arxiv.org/pdf/1409.1556.pdf.
    [17] Szegedy C, Liu W, Jia Y, et al. Going deeper with convolutions[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. Washington D C: IEEE Computer Society Press, 2015: 1−9.
    [18] He K M, Zhang X Y, Ren S Q, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE conference on computer vision and pattern recognition, Vegas: IEEE, 2016: 770−778.
    [19] 李彦冬, 郝宗波, 雷航. 卷积神经网络研究综述[J]. 计算机应用, 2016, 36(9):2508−2515, 2565. doi: 10.11772/j.issn.1001-9081.2016.09.2508

    Li Y D, Hao Z B, Lei H. A review of convolutional neural networks[J]. Computer applications, 2016, 36(9): 2508−2515, 2565. doi: 10.11772/j.issn.1001-9081.2016.09.2508
    [20] Zeiler M D, Fergus R. Stochastic pooling for regularization of deep convolutional neural networks[C/OL]. arXiv, 2013 [2018−04−16]. https://arxiv.org/pdf/1301.3557.pdf.
    [21] Nair V, Hinton G E, Farabet C. Rectified linear units improve restricted Boltzmannmachines[C]//Processing of the 27th International Conference on Machine Learning. Haifa: International Machine Learning Society (IMLS), 2010: 807−714.
    [22] 周飞燕, 金林鹏, 董军. 卷积神经网络研究综述[J]. 计算机学报, 2017, 40(6):1229−1251. doi: 10.11897/SP.J.1016.2017.01229

    Zhou F Y, Jin L P, Dong J. Summary of convolution neural network research[J]. Acta Computer Science, 2017, 40(6): 1229−1251. doi: 10.11897/SP.J.1016.2017.01229
    [23] 刘涛, 周先春, 严锡君. 多通道多模式融合LBP特征的纹理相似度计算[J]. 计算机应用研究, 2018, 35(12):3803−3806. doi: 10.3969/j.issn.1001-3695.2018.12.063

    Liu T, Zhou X C, Yan X J. Computation of texture similarity based on multi-channel and multi-mode LBP features[J]. Computer Applications, 2018, 35(12): 3803−3806. doi: 10.3969/j.issn.1001-3695.2018.12.063
    [24] 尚俊. 基于HOG特征的目标识别算法研究[D]. 武汉: 华中科技大学, 2012.

    Shang J. Target recognition algorithm based on HOG features[D]. Wuhan: Huazhong University of Science and Technology, 2012.
    [25] 张盼. 基于混淆矩阵的分类器选择集成方法研究[D]. 焦作: 河南理工大学, 2016.

    Zhang P. Ensemble method of classifier selection based on confusion matrix[D]. Jiaozuo: Henan Polytechnic University, 2016.
  • 加载中
图(9) / 表(4)
计量
  • 文章访问数:  2939
  • HTML全文浏览量:  1238
  • PDF下载量:  184
  • 被引次数: 0
出版历程
  • 收稿日期:  2018-11-12
  • 修回日期:  2019-03-08
  • 网络出版日期:  2019-10-01
  • 刊出日期:  2019-11-01

目录

    /

    返回文章
    返回