一种基于改进YOLO v3的小目标检测方法  

A Small Target Detection Method based on Improved YOLO v3

作　　者：蔡鸿峰 吴观茂[1] CAI Hongfeng;WU Guanmao(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan Anhui 232001)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《湖北理工学院学报》2021年第2期33-36,47,共5页Journal of Hubei Polytechnic University

基　　金：安徽省自然科学基金面上项目(项目编号:1908085MF189)。

摘　　要：为了提升小目标物体检测的精度,尽可能避免发生漏检情况,提出了一种改进的YOLO v3算法。选用Darknet-49为主干网络,将原有的3个检测尺度扩展为5个。同时,引入DIoU函数,对损失函数进行了改进,将PASCAL VOC 2012数据集作为测试数据集和训练数据集进行实验。研究发现,改进后的YOLO v3算法与原YOLO v3算法相比,mAP值提升了约2.4%,且检测速度与原算法接近。实验结果表明,改进原YOLO v3算法网络结构和损失函数,可以增强算法对小目标物体的检测精度。In order to improve the accuracy of small object detection and reduce the occurrence of missed detection,an improved YOLO v3 algorithm is proposed.The algorithm uses Darknet-49 as the backbone network,and expands the original three detection scales to five detection scales.At the same time,DIoU function is introduced to improve the loss function.PASCAL VOC 2012 data set is used as test data set and training data set for experiments.Compared with the original YOLO v3 algorithm,the map value of the improved YOLO v3 algorithm has been improved by about 2.4%,and the detection speed is close to the original algorithm.The experimental results show that by improving the network structure and loss function of the original YOLO v3 algorithm,the detection accuracy of small objects can be enhanced.

关 键 词：小目标 特征融合 损失函数 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]