高级检索
    朱德利, 林智健. 基于MF-SSD卷积神经网络的玉米穗丝目标检测方法[J]. 华南农业大学学报, 2020, 41(6): 109-118. doi: 10.7671/j.issn.1001-411X.202006025
    引用本文: 朱德利, 林智健. 基于MF-SSD卷积神经网络的玉米穗丝目标检测方法[J]. 华南农业大学学报, 2020, 41(6): 109-118. doi: 10.7671/j.issn.1001-411X.202006025
    shu
    ZHU Deli, LIN Zhijian. A corn silk detection method based on MF-SSD convolutional neural network[J]. Journal of South China Agricultural University, 2020, 41(6): 109-118. doi: 10.7671/j.issn.1001-411X.202006025
    Citation: ZHU Deli, LIN Zhijian. A corn silk detection method based on MF-SSD convolutional neural network[J]. Journal of South China Agricultural University, 2020, 41(6): 109-118. doi: 10.7671/j.issn.1001-411X.202006025
    shu

    基于MF-SSD卷积神经网络的玉米穗丝目标检测方法

    A corn silk detection method based on MF-SSD convolutional neural network

      摘要
    • 摘要:
      目的  玉米穗丝是玉米的授粉器官，生长发育状况会影响玉米的产量。为了在玉米生长状态监测和产量预测工作中实时准确识别玉米穗丝，提出一种基于多特征融合SSD (MF-SSD)卷积神经网络的玉米穗丝检测模型。
      方法  基于特征图对玉米穗丝进行检测，在VGG16-SSD的基础上，用MobileNet替换特征提取器，加入多层特征融合结构，得到MF-SSD网络模型；通过网络优化调整，试验了MF-SSD-cut-3、MF-SSD和MF-SSD-add-3共3种网络结构，优选出检测性能最好的网络结构用于玉米穗丝检测。基于玉米穗丝图像数据集，应用0~180°随机旋转原始图像和水平翻转、平移原始图像2种数据增广技术提升模型训练效果。对是否使用二次训练策略和是否使用Focal loss解决样本不平衡问题进行了试验，并对比分析Loss的下降过程。
      结果  通过加入多层特征融合结构对SSD模型改进后能够提高网络的检测能力，提升识别速度。与VGG16-SSD相比，MF-SSD在交并比指标方面的平均精度提高7.2%，对玉米穗丝小目标检测的平均召回率提高19.6%，检测速度最高能提升18.7%。在存储空间和运行时间有较高要求的嵌入式环境下，MF-SSD-cut-3模型在满足检测效果的前提下，以较小的空间代价获得了相对较短的运行时间；在不考虑空间和时间因素的情况下，MF-SSD模型获得更好的检测效果。二次训练策略提高了网络的收敛速度和模型的稳定性；Focal loss有效解决了SSD算法中正负样本数量不平衡问题，使网络模型的训练更容易收敛。
      结论  MF-SSD模型对小目标的检测能力能满足农业生产中对玉米穗丝的实时检测需要，可以用于玉米生长状态的自动监控和产量的精准预测。

       

      Abstract:
      Objective  Corn silk is the pollination organ of maize, its growth state will affect the yield of corn. In order to identify the corn silk in real-time and accurately in corn growth state monitoring and yield prediction, a corn silk detection model based on multi-feature fusion SSD (MF-SSD) convolutional neural network was proposed.
      Method  Corn silk detection was based on feature images. The MF-SSD network model was modified from VGG16-SSD through replacing feature extractor by MobileNet and integrating multi-layer feature fusion structure. By optimizing and adjusting the network, three kinds of MF-SSD with different network structures (MF-SSD-cut-3, MF-SSD and MF-SSD-add-3) were tested, and the structure with the best detection performance was selected for corn silk detection. Based on the image data set of corn silk, two kinds of data augmentation techniques (randomly rotating original image from 0 to 180°, horizontally rolling over or translating original image) were applied to improve the training effect of the model. Whether using secondary training strategy and Focal loss to solve sample number imbalance was investigated, and the decrease process of Loss was compared and analyzed.
      Result  The improved SSD model added with multi-layer feature fusion structure could improve the detection ability and recognition speed of network. Compared with VGG16-SSD, the average accuracy of intersection over union increased by 7.2%, the average recall of small target detection of corn silk increased by 19.6%, and the detection speed increased by 18.7%. In the embedded environment having high demand for storage space and run time, MF-SSD-cut-3 obtained shorter run time with smaller storage space in the premise of satisfying detection effect. MF-SSD obtained better detection effect in the condition of taking no account of storage space and run time. The secondary training strategy improved the network convergence speed and model stability. Focal loss effectively solved number imbalance problem of positive and negative samples, and made the training of network model more convergent.
      Conclusion  The detection effect of the proposed MF-SSD model for small targets can meet the needs of real-time detection of corn silk in agricultural production. It can be used for automatic monitoring of corn growth state and yield prediction.

       

      • HTML全文
      • 参考文献(22)
      • 相关文章
      • 施引文献
    • 资源附件(0)

    /

    返回文章
    返回