欢迎您访问《智慧农业(中英文)》官方网站! English

Smart Agriculture ›› 2021, Vol. 3 ›› Issue (1): 63-74.doi: 10.12133/j.smartag.2021.3.1.202102-SA066

• • 上一篇    下一篇

基于YOLOv4和自适应锚框调整的谷穗检测方法

郝王丽1(), 尉培岩1, 郝飞2, 韩猛1, 韩冀皖1, 孙玮蓉1, 李富忠1()   

  1. 1.山西农业大学 软件学院,山西 晋中,030801
    2.陕西师范大学 计算机学院,陕西 西安,710119
  • 收稿日期:2021-02-25 修回日期:2021-03-26 出版日期:2021-03-30
  • 基金资助:
    山西省高等教育创新工程 (2020L0154)
  • 作者简介:郝王丽, E-mail: hanmeng10@126.com
  • 通信作者: 李富忠, E-mail: 

Foxtail Millet Ear Detection Approach Based on YOLOv4 and Adaptive Anchor Box Adjustment

HAO Wangli1(), YU Peiyan1, HAO Fei2, HAN Meng1, HAN Jiwan1, SUN Weirong1, LI Fuzhong1()   

  1. 1.School of Software, Shanxi Agricultural University, Shanxi 030801, China
    2.School of Computer Science, Shaanxi Normal University, Xi'an 710119, China
  • Received:2021-02-25 Revised:2021-03-26 Online:2021-03-30
  • corresponding author: LI Fuzhong(1969-), male, Ph.D., professor, research interest is smart agriculture. Tel: 0354-6287093. E-mail: 
  • About author:HAO Wangli(1988-), female, Ph.D., lecturer, research interests is artificial and smart agriculture. E-mail: hanmeng10@126.com
  • Supported by:
    Shanxi Province Higher Education Innovation Project of China (2020L0154)

摘要:

Foxtail millet ear detection and counting are essential for the estimation of foxtail millet production and breeding. However, traditional foxtail millet ear counting approaches based on manual statistics are usually time-consuming and labor-intensive. In order to count the foxtail millet ears accurately and efficiently, an adaptive anchor box adjustment foxtail millet ear detection method was proposed in this research. Ear detection dataset was firstly established, including 784 images and 10,000 ear samples. Furthermore, a novel foxtail millet ear detection approach based on YOLOv4 (You Only Look Once) was developed to quickly and accurately detect the ear of foxtail millet in the specific box. For verifying the effectiveness of the proposed approach, several criteria, including the mean average Precision, F1-score,Recall and mAP were employed. Moreover, ablation studies were designed to validate the effectiveness of the proposed method, including (1) evaluating the performance of the proposed model through comparing with other models (YOLOv2, YOLOv3 and Faster-RCNN); (2) evaluating the model with different Intersection over Union (IOU) thresholds to achieve the optimal IOU thresholds; (3) evaluating the foxtail millet ear detection with or without anchor boxes adjustment to verify the effectiveness of the adjustment of anchor boxes;(4) evaluating the changing reasons of model criteria and (5) evaluating the foxtail millet ear detection with different input original image size respectively. Experimental results showed that YOLOv4 could obtain the superior ear detection performance. Specifically, mAP and F1-score of YOLOv4 achieved 78.99% and 83.00%, respectively. The Precision was 87% and the Recall was 79.00%, which was about 8% better than YOLOv2, YOLOv3 and Faster RCNN models, in terms of all criteria. Moreover, experimental results indicates that the proposed method is superior with promising accuracy and faster speed.

关键词: foxtail millet ear detection, YOLOv4, deep neural network, dataset, adaptive anchor box adjustment

Abstract:

谷穗的检测和计数对于预测谷子产量和育种至关重要。但是,传统的谷穗计数主要基于人工统计,既费时又费力。为解决上述问题,本研究首先建立了一个包含784张图像和10,000个谷穗样本的谷穗检测数据集。提出了一种基于YOLOv4和自适应锚框调整的谷穗检测方法,可快速准确地检测特定框中的谷穗。通过自适应地调整锚框,可生成符合谷穗目标的候选框,从而提升检测的准确率。为验证该方法的有效性,采用了多个标准,包括平均精度(mAP),F1得分(F1-Score),精度(Precision)和召回率(Recall)进行评价。此外,设计了对比试验验证所提出方法的有效性,包括与其他模型(YOLOv2,YOLOv3和Faster-RCNN)进行比较来评估模型的性能,评估模型在不同交并比(IOU)取值下的性能,评估模型在自适应锚框调整下的谷穗检测性能,评估引起模型评价标准变化的原因,以及评估模型在不同原始输入图像尺寸下的性能。试验结果表明,YOLOv4获得了良好的谷穗检测性能。YOLOv4的mAP达到78.99%,F1-score达到83.00%,Precision达到87%和Recall达到79.00%,在所有评价标准上均比其他比较模型高出8%。试验结果表明,该方法具有较好的准确性和高效性。

Key words: 谷穗检测, YOLOv4, 深度神经网络, 数据集, 自适应锚框调整

中图分类号: