融合非对称卷积的遥感图像目标检测算法  

Object detection algorithm of remote sensing image based on asymmetric convolution

作　　者：齐梦林 陈炳才 张繁盛[2] 潘旭 彭相澍 Qi Menglin;Chen Bingcai;Zhang Fansheng;Pan Xu;Peng Xiangshu(School of Computer Science and Technology,Xinjiang Normal University,Urumqi 830054,China;College of Computer Science and Technology,Dalian University of Technology,Dalian 116024,China)

机构地区：[1]新疆师范大学计算机科学技术学院,乌鲁木齐830054 [2]大连理工大学计算机科学与技术学院,大连116024

出　　处：《电子测量技术》2023年第7期125-132,共8页Electronic Measurement Technology

基　　金：国家自然科学基金(61961040,61771089);新疆维吾尔自治区“天山青年计划”(2018Q024);新疆维吾尔自治区区域协同创新专项(科技援疆计划)(2020E0247,2019E0214)资助。

摘　　要：遥感图像中的目标具有背景复杂、方向多变等特点。利用传统方法进行遥感图像目标检测过程复杂且费时,存在精度低,漏检率高等问题。针对以上问题,提出一种改进的YOLOv5-AC算法,该算法以YOLOv5s模型为基础,首先在原有的Backbone中构建非对称卷积结构,增强模型对翻转和旋转目标的鲁棒性;其次在主干网络的C3模块中引入坐标注意力机制提升特征提取能力,并使用Acon自适应激活函数激活;最后使用CIOU作为定位损失函数以提升模型定位精度。改进后的YOLOv5-AC模型在NWPU VHR-10和RSOD数据集上进行实验,平均精确度均值分别达到了94.0%和94.5%,分别比原版YOLOv5s提升了1.8%和2.3%,有效提高了遥感图像目标检测精确度。The object of remote sensing image has the characteristics of complex background and changeable direction.The process of object detection in remote sensing image using traditional methods is complex and time-consuming,with low accuracy and high rate of missed detection.To solve the above problems,we propose an improved YOLOv5-AC algorithm.This algorithm bases on the YOLOv5s model.First,an asymmetric convolution structure is built in the original Backbone to enhance the robustness of the model to flipped and rotated targets;Secondly,coordinate attention mechanism is introduced into C3 module of backbone network to improve feature extraction capability,and Acon(Activate Or Not) adaptive activation function is used for activation;Finally,we use CIOU as the location loss function to improve the positioning accuracy of the model.The improved YOLOv5-AC model was tested on NWPU VHR-10 and RSOD datasets,and the average accuracy reached 94.0% and 94.5%,respectively,1.8% and 2.3% higher than the original YOLOv5s,which effectively improved the object detection accuracy of remote sensing images.

关 键 词：遥感图像 目标检测 YOLOv5 非对称卷积 注意力机制 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]