基于Ghost-CA-YOLOv4的果园障碍物检测算法  

Orchard Obstacle Detection Algorithm Based on Ghost-CA-YOLOv4

作　　者：景亮[1] 鲍致远 JING Liang;BAO Zhiyuan(School of Electrical and Information Engineering,Jiangsu University,Zhenjiang 212013,China)

机构地区：[1]江苏大学电气信息工程学院,江苏镇江212013

出　　处：《软件导刊》2023年第11期180-185,共6页Software Guide

摘　　要：针对无人农机在复杂果园环境中检测障碍物存在实时性、精确性不足的问题,提出基于YOLOv4改进的Ghost-CA-YOLOv4果园障碍物检测算法。选用轻量化网络GhostNet作为主干特征提取网络,在GhostNet中引入注意力机制(CA)对空间信息和通道信息进行融合,提升目标关注度。同时,使用Soft-CIoU-NMS算法增强在目标密集或重叠时的检测效果。自制的果园障碍物数据集实验表明,所提算法的mAP值达到90.75%,在检测精度保持不变的情况下,模型大小压缩为传统YOLOv4模型的18.24%,FPS为45,可更好地为无人农机作业提供安全保障。Aiming at the problem that the detection of obstacles by unmanned agricultural machinery in complex orchard environment is not real-time and accurate,an improved Ghost-CA-YOLOv4 orchard obstacle detection algorithm based on YOLOv4 is proposed.This algorithm selects the lightweight network GhostNet as the backbone feature extraction network,and introduces CA mechanism in GhostNet to fuse spatial information and channel information,so as to improve the target attention.Soft-CIoU-NMS algorithm is used to enhance the detection effect when the targets are dense or overlapping.To verify the effectiveness of the algorithm,the mAP value on the self-made orchard obstacle data set reached 90.75%after training in the Python framework.With the detection accuracy unchanged,the model size is compressed to 18.24%of the original YOLOv4,and the FPS is 45,which can better provide security for the operation of unmanned agricultural machinery.

关 键 词：深度学习 目标检测 GhostNet 注意力机制 Soft-CIoU-NMS 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] TP183[自动化与计算机技术—控制科学与工程]