改进YOLOv3的多尺度高分辨率特征增强图像目标检测  

Multi-scale High-resolution Feature Enhancement Image Target Detection Based on Improved YOLOv3

作　　者：杨文姬[1] 李浩 王映龙[3] 梅梦 YANG Wen-ji;LI Hao;WANG Ying-long;MEI Meng(School of Software,Jiangxi Agricultural University,Nanchang 330045,China;Key Laboratory of Agricultural Information Technology,Jiangxi Agricultural University,Nanchang 330045,China;School of Computer and Information Engineering,Jiangxi Agricultural University,Nanchang 330045,China;Business School of Jiangxi Province,Nanchang 330100,China)

机构地区：[1]江西农业大学软件学院,南昌330045 [2]江西农业大学农业信息技术重点实验室,南昌330045 [3]江西农业大学计算机与信息工程学院,南昌330045 [4]江西省商务学校,南昌330100

出　　处：《小型微型计算机系统》2023年第6期1311-1317,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61462038)资助;江西省自然科学基金项目(20212BAB212005)资助;江西省教育厅科技项目(GJJ190217)资助;浙江大学国家重点实验室开放课题项目(A2029)资助.

摘　　要：针对目标检测存在背景复杂,遮挡严重,尺度大小不一的问题,本文基于YOLOv3进行改进,提出了一种多尺度高分辨率的特征融合网络YOLOv3-F.首先,在backbone和neck之间添加空间金字塔池化结构(Spatial Pyramid Pooling),特征图经过局部特征与全局特征相融合后,丰富了特征图的表达能力,有利于待检测图像中目标大小差异较大的情况;其次,三尺度检测层增加为四尺度检测层,通过高分辨率特征图位置信息的保留,提高小目标识别的精度.最后,将YOLOv3损失函数改为CIoU,提高目标检测框回归的效率,降低目标漏检率.将改进算法在自制数据集COCO-CT6上进行试验,误检率降低1.8%,精度提高了3.6%.Aiming at the problems of complex background,severe occlusion,and different scales in target detection,this paper improves on YOLOv3 and proposes a multi-scale and high-resolution feature fusion network YOLOv3-F.Firstly,Spatial Pyramid Pooling is added between backbone and neck.After the feature map is fused with local features and global features,it enriches the expressive ability of the feature map,which is conducive to the situation where the target size in the image to be detected is greatly different.Secondly,the three-scale detection layer is increased to a four-scale detection layer,and the accuracy of small target recognition is improved by retaining the position information of the high-resolution feature map.Finally,the YOLOv3 loss function is changed to CIoU to improve the efficiency of the target detection box regression and reduce the target miss rate.The improved algorithm is tested on the self-made data set COCO-CT6,the false detection rate is reduced by 1.8%,and the accuracy is increased by 3.6%.

关 键 词：图像目标检测 多尺度特征 YOLOv3 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]