计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (4): 221-228.DOI: 10.3778/j.issn.1002-8331.2106-0239

• 图形图像处理 • 上一篇    下一篇

基于改进YOLOv4的植物叶茎交点目标检测研究

李文婧,徐国伟,孔维刚,郭风祥,宋庆增   

  1. 1.天津工业大学 电气与电子工程学院,天津 300387
    2.天津工业大学 计算机科学与技术学院,天津 300387
  • 出版日期:2022-02-15 发布日期:2022-02-15

Research on Target Detection of Plant Leaf-Stem Intersection Based on Improved YOLOv4

LI Wenjing, XU Guowei, KONG Weigang, GUO Fengxiang, SONG Qingzeng   

  1. 1.School of Electrical and Electronic Engineering, Tiangong University, Tianjin 300387, China
    2.School of Computer Science and Technology, Tiangong University, Tianjin 300387, China
  • Online:2022-02-15 Published:2022-02-15

摘要: 由于植物根茎交点目标较小,识别率低,且在使用嵌入式设备进行植物移植与栽培的过程中资源及功耗受限。针对这类问题,提出了一种基于改进YOLOv4的目标检测解决方法,并设置于本场景。采集8?629张植物叶茎数据集图像,并对这些植物叶茎数据集进行标注,利用生成对抗网络(generative adversarial network,GAN)进行数据增强预处理。改进YOLOv4目标检测算法,选取4个不同尺度的锚框,以获得更多植物叶茎交点信息,同时对网络的结构和损失函数进行局部优化,使得网络在训练过程中更易于拟合目标。将主干网络更改为GhostNet网络,大幅度减少参数量与计算量,更利于在移动设备上的轻量化模型部署。实验结果表明,优化后的YOLOv4-GhostNet轻量化网络在保证检测精度的前提下,检测速度提高到79.3?frame/s,较YOLOv4提高了36.45%,网络帧率提高了51.07%,模型参数量减小了36.06%,能够有效检测叶茎交点目标。

关键词: 目标检测, 数据增强, 对抗网络, YOLOv4, GhostNet

Abstract: Due to the small target of the plant roots and stems, the recognition rate is low, and the resources and power consumption are limited in the process of plant transplantation and cultivation using embedded devices. Aiming at this kind of problem, an improved YOLOv4 target detection solution method is proposed, which is more equipped in this scene. Firstly, 8629 images of the plant leaf and stem data sets are collected, and these plant leaf and stem data sets are labeled, and the generative adversarial network(GAN) is used for data enhancement preprocessing. Then, the YOLOv4 target detection algorithm is improved, four anchor boxes of different scales are selected to obtain more plant leaf-stem intersection information, and the structure and loss function of the network are locally optimized to make the network easier fit the target during the training process. Finally, the backbone network is changed to the GhostNet network, which greatly reduces the amount of parameters and calculations, and is more conducive to the deployment of lightweight models on mobile devices. The experimental results show that the optimized YOLOv4-GhostNet lightweight network can increase the detection speed to 79.3 frame/s under the premise of ensuring the detection accuracy, which is 36.45% higher than YOLOv4, the network frame rate is increased by 51.07%, and the amount of model parameters is reduced by 36.06%, can effectively detect the leaf-stem intersection target.

Key words: target detection, data enhancement, confrontation network, YOLOv4, GhostNet