高级检索+

基于轻量化改进模型的小麦白粉病检测装置研发

Development of Detector for Wheat Powdery Mildew Based on Lightweight Improved Deep Learning Model

  • 摘要: 为快速、全面的监测大田小麦病害,并结合小麦发病特征实现对小麦不同生长部位的病害进行识别,设计了一款便携式小麦白粉病病害检测装置,其由双相机采集模块和主控模块组成,配合病害检测软件系统实现对小麦多部位的白粉病害采集与检测。为保证模型在检测装置部署的可行性,提出了一种基于YOLO v7-tiny模型轻量化改进的白粉病目标检测模型(YOLO v7tiny-ShuffleNet v1,YT-SFNet)。为验证该轻量化模型的准确率和检测速度,与YOLO v7-tiny模型进行训练对比,结果表明YT-SFNet模型相较于YOLO v7-tiny在平均精度上提高了0.57个百分点;在检测时间和模型内存占用量上分别下降了2.4 ms和3.2 MB。最后将轻量化模型和软件系统移植至装置主控模块,制作测试集对装置的检测准确率和检测速度进行了性能测试。其对于测试集的识别准确率为86.2%,检测速度上有较好的稳定性,且单幅病害图像从处理、检测及显示保存的全过程平均耗时为0.507 9 s。

     

    Abstract: Wheat diseases have frequently threatened the yield and quality of wheat production. In order to quickly and comprehensively monitor wheat diseases in the field and identify diseases in different growth parts of wheat based on the characteristics of wheat disease, a dual camera wheat disease detection device based on a lightweight model was designed. The device was composed of a dual camera acquisition module and a main control module, and it can collect and detect wheat powdery mildew at multiple locations in cooperation with the disease detection software system. In order to ensure the feasibility of the model deployment in the detection device, a lightweight improved powdery mildew target detection model based on YOLO v7-tiny model(YOLO v7tiny-ShuffleNet v1, YT-SFNet) was proposed. To verify the accuracy and detection speed of the lightweight model, it was trained and compared with the YOLO v7-tiny model. The results showed that the YT-SFNet model improved the average accuracy by 0.57 percentage points compared with YOLO v7-tiny model. The detection time and model size were decreased by 2.4 ms and 3.2 MB, respectively. Finally, the lightweight model and software system were transplanted to the main control module of the device, and a test set was created to test the performance of the device’s detection accuracy and detection speed. Its recognition accuracy for the test set was 86.2%, with good stability in detection speed, and the average time spent on the entire process of processing, detecting, and displaying and saving a single disease image was 0.507 9 s.

     

/

返回文章
返回