自然环境下基于增强YOLOv3的百香果目标检测  

Target Detection of Passiflora edulis Sims.Based on Enhanced YOLOv3 under Natural Environment

作　　者：张展榜 罗志聪[1] 周志斌 李鹏博 孙奇燕[2] ZHANG Zhan-bang;LUO Zhi-cong;ZHOU Zhi-bin(College of Mechanical and Electronic Engineering,Fujian Agriculture and Forestry University,Fuzhou,Fujian 350100)

机构地区：[1]福建农林大学机电工程学院,福建福州350100 [2]福建农林大学计算机与信息学院,福建福州350002

出　　处：《安徽农业科学》2022年第6期186-192,197,共8页Journal of Anhui Agricultural Sciences

基　　金：海峡博士后交流资助计划项目。

摘　　要：为解决当前流行的目标检测模型对自然环境下百香果由于目标密集互相遮挡所致的检测效率低等问题,以YOLOv3网络为基础,提出了一种基于增强的YOLOv3百香果目标检测算法。首先,针对百香果目标尺寸的特点,利用以交并比为距离度量的改进K-means++算法,重新获取与目标果实相匹配的锚选框,提高对目标的框选精度以及模型的收敛速度;其次,在输出网络中将用来筛选目标预测框的Soft-NMS算法通过线性函数的形式对其高斯函数的抑制参数进行改进,以提高模型在不同密集场景下的适应性和检测能力;最后,利用增强的YOLOv3模型在经过预处理后的百香果数据集上进行多次试验对比,结果表明增强后的YOLOv3目标检测算法平均精度均值(mAP)达到94.62%,F_(1)值达到94.34%,较原YOLOv3算法分别提升了4.58和3.68百分点,平均检测速度为25.45帧/s,基本满足了自然环境下百香果目标检测的精准性和实时性要求。In order to solve the problem of low detection efficiency caused by the dense mutual occlusion of Passiflora edulis Sims.under the natural environment by the current popular target detection model,we proposed an enhanced YOLOv3 target detection algorithm of P.edulis based on the YOLOv3 network.First of all,according to the target size characteristics of P.edulis,the improved K-means++algorithm based on the intersection over union as the distance metric was used to retrieve the anchor box matching the target fruit,improve the accuracy of the target box selection and the convergence speed of the model.Secondly,in the output network,the Soft-NMS algorithm for was used to filter target prediction box,improve the suppression parameters of its Gaussian function in the form of linear function,so as to improve the adaptability and detection ability of the model in different dense scenarios.Finally,the enhanced YOLOv3 model was used to make multiple comparative experiments on the data set of P.edulis after pre-treatment.The experimental results showed that the mean average precision(mAP)of the enhanced YOLOv3 target detection algorithm reached 94.62%,F_(1) reached 94.34%,which increased 4.58 and 3.68 percentage points respectively,and the average detection speed was 25.45 frames per second,which basically met the accuracy and real-time requirements of target detection of P.edulis under natural environments.

关 键 词：百香果 自然场景 密集目标检测 增强YOLOv3 非极大抑制算法 

分 类 号：S126[农业科学—农业基础科学]