基于改进SSD算法的轻量化三七病害检测

    Lightweight detection of Panax notoginseng disease based on improved SSD algorithm

    • 摘要:
      目的  针对目前三七Panax notoginseng病害识别模型结构复杂、参数庞大,难以实现在移动设备上部署的问题,提出一种基于SSD(Single shot multibox detector)目标检测的改进模型,以期实现三七病害检测的便捷化、快速化与精准化。
      方法  基于SSD模型架构,采用轻量化卷积神经网络(MobileNet)替换原始特征提取网络(VGG16),降低主干网络的参数量与计算量,同时根据人类视觉皮层中群智感受野(pRF)的大小与其视网膜图中偏心率之间的函数关系,构建RFB模块,用该模块替换原SSD模型框架顶部卷积层,从而增强网络深层特征,提高轻量化模型的检测精度与检测速度,实现多尺度三七病害检测。
      结果  与SSD模型相比,RFB-MobileNet-SSD模型网络参数量和参数计算量分别降低了96.67%和96.10%。在不同天气条件下应用模型对4种不同病害数据进行验证发现,改进模型的准确率提高了4.6个百分点,召回率提高了6.1个百分点,F1精度提高了5.4个百分点,单幅图像检测时间由SSD模型的0.073 s缩短为0.020 s,尺寸仅为SSD模型的54.6%。
      结论  改进后模型不仅能够满足三七叶片病害实时检测的要求,且更利于在移动设备中部署。此外,RFB-MobileNet-SSD对于小区域病害检测能力表现更优且在复杂环境下抗干扰能力更强,更适合田间环境下的三七病害检测。

       

      Abstract:
      Objective  Aiming to address the problem of the current Panax notoginseng disease identification model, with complex structure and large number of parameters, hindering deployment on mobile devices, an improved model based on single shot multibox detector (SSD) target detection is proposed to enable convenient, fast and accurate P. notoginseng disease detection.
      Method  Based on the SSD model architecture, the original feature extraction network (VGG16) was replaced by a lightweight convolutional neural network (MobileNet) to reduce the number of parameters and computation amount of the backbone network. Meanwhile, the RFB module was constructed based on the functional relationship between the size of the population wise receptive field (pRF) in human visual cortex and its eccentricity in the retinogram. The top convolutional layer of the original SSD model framework was replaced by the RFB module to enhance the deep features of the network, improve the detection accuracy and detection speed of the lightweight model, and enable multi-scale P. notoginseng disease detection.
      Result  Compared with the SSD model, the RFB-MobileNet-SSD model reduced the number of network parameters and the computation amount by 96.67% and 96.10% respectively. The model validation using four different disease data under different weather conditions revealed that the improved model improved the accuracy by 4.6 percentage point, recall by 6.1 percentage point, F1 accuracy by 5.4 percentage point, and the time of single image detection was shortened from 0.073 s of the SSD model to 0.020 s, and the size was only 54.6% of the SSD model.
      Conclusion  The improved model not only meets the purpose of real-time detection ofP. notoginseng leaf diseases, but is also more convenient for deployment in mobile devices. Moreover, RFB-MobileNet-SSD shows improved performance for small area disease detection and is more resistant to interference in complex environments, making it more suitable for P. notoginseng disease detection in field environment.

       

    /

    返回文章
    返回