高级检索
    胡静, 陈志泊, 杨猛, 张荣国, 崔亚稷. 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. DOI: 10.13332/j.1000-1522.20180007
    引用本文: 胡静, 陈志泊, 杨猛, 张荣国, 崔亚稷. 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. DOI: 10.13332/j.1000-1522.20180007
    Hu Jing, Chen Zhibo, Yang Meng, Zhang Rongguo, Cui Yaji. Plant leaf segmentation method based on fully convolutional neural network[J]. Journal of Beijing Forestry University, 2018, 40(11): 131-136. DOI: 10.13332/j.1000-1522.20180007
    Citation: Hu Jing, Chen Zhibo, Yang Meng, Zhang Rongguo, Cui Yaji. Plant leaf segmentation method based on fully convolutional neural network[J]. Journal of Beijing Forestry University, 2018, 40(11): 131-136. DOI: 10.13332/j.1000-1522.20180007

    基于全卷积神经网络的植物叶片分割算法

    Plant leaf segmentation method based on fully convolutional neural network

    • 摘要:
      目的植物叶片分割旨在从背景中分割出叶片区域,去除背景对象干扰。这对植物病害识别和物种鉴定具有重大意义。
      方法本文设计了基于全卷积神经网络的植物叶片分割算法。首先,目标函数用对数逻辑函数代替复杂的Softmax多类预测函数,从而将分割任务转化为适合于植物叶片分割的二分类问题;其次,把批归一化技术引入全卷积神经网络,从而改善网络整体的收敛性。最后,针对当前植物叶片分割研究中缺乏评估指标的状况,设计了新的评估协议——受试者工作特征曲线,该曲线反映了不同阈值情况下植物叶片图像分割的召回率与误报率之间的变化情况。
      结果本文提出的算法降低了全卷积神经网络的参数复杂度,改善了网络的收敛性。实验结果表明,该方法比Leafsnap提到的基于颜色的分割方法更完整地分割了植物叶片区域;提出的ROC曲线能够充分评估植物叶片的分割性能。
      结论与传统方法相比,基于深度学习的植物叶片分割方法实现了输入图像的端对端处理,无需图像转换、噪声滤波和形态运算等预处理技术,因此在植物叶片分割上具有可行性。

       

      Abstract:
      ObjectivePlant leaf segmentation aims to segment leaf regions from backgrounds for removing background object interferes, which is important for plant disease recognition and species identification.
      MethodIn this paper, a fully convolutional neural network (FCNN) was designed for plant leaf image segmentation. First, a log-logic function as objective function replaces the complex Softmax function, which transforms the segmentation task into a binary classification problem suitable for plant leaf segmentation. Second, the batch normalization (BN) technology was introduced into the FCNN, which improved the convergence of the whole FCNN. Finally, due to the lack of evaluation index in the research of plant leaf segmentation, receiver operating characteristic (ROC) curve, as a new evaluation protocol, was designed. It reflected the changes between recall rate and false alarm rate of plant leaf segmentation under different threshold settings.
      ResultThe method reduced the complexity of parameters and improved the convergence performance of FCNN. Experimental results showed that this method was more complete to segment the leaf image than the color-based method in Leafsnap. The proposed ROC curve adequately evaluated the performance of plant leaf segmentation.
      ConclusionCompared with traditional plant leaf segmentation methods, the proposed method based on deep learning realizes input image by end to end processing, and does not require pre-processing like image conversion, noise filter and morphological operations, etc. Therefore, this method can be used for leaf segmentation.

       

    /

    返回文章
    返回