改进YOLOv4的遥感图像目标检测算法  

Improved YOLOv4 algorithm for remote sensing image object detection

作　　者：闵锋[1,2] 况永刚 毛一新 彭伟明 郝琳琳 MIN Feng;KUANG Yong-gang;MAO Yi-xin;PENG Wei-ming;HAO Lin-lin(School of Computer Science and Engineering,Wuhan Institute of Technology,Wuhan 430205,China;Hubei Province Key Laboratory of Intelligent Robot,Wuhan Institute of Technology,Wuhan 430205,China)

机构地区：[1]武汉工程大学计算机科学与工程学院,湖北武汉430205 [2]武汉工程大学智能机器人湖北省重点实验室,湖北武汉430205

出　　处：《计算机工程与设计》2024年第2期396-404,共9页Computer Engineering and Design

基　　金：国家自然科学基金项目(62171328)。

摘　　要：为有效解决遥感图像目标检测算法在复杂背景下的检测效果不佳的问题,提出一种改进YOLOv4的目标检测算法。设计一种跨阶段残差结构,替换原主干网络的简单残差结构,降低模型参数量和计算负担;引入CBAM注意力机制,加强CSP模块间有效特征交互;使用跨阶段分层卷积模块重构特征融合阶段对深层特征图的处理方式,防止网络退化和梯度消失;采用Mish激活函数,增强融合网络对非线性特征的提取能力。在RSOD、DIOR数据集上的实验结果表明,改进YOLOv4算法的测试mAP相比原YOLOv4算法分别高出4.5%、7.3%,其检测速度分别达到48 fps、45 fps,在保证实时性的同时检测精度有较大提升。To effectively solve the problem of poor detection effects of remote sensing image target detection algorithm in complex background,an improved YOLOv4 target detection algorithm was proposed.A cross-stage residual structure was designed to replace the simple residual structure in the CSP module of the original backbone network to reduce the amount of model parameters and computational burden.The CBAM attention mechanism was introduced to strengthen the interaction of effective features between CSP modules.The cross-stage hierarchical convolution module was used to reconstruct the processing method of the deep feature map in the feature fusion stage to prevent network degradation and gradient disappearance.The Mish activation function was used to enhance the extraction ability of the fusion network for nonlinear features.Experimental results on the RSOD and DIOR data sets show that the test mAP of the improved YOLOv4 algorithm is 4.5%and 7.3%higher than that of the original YOLOv4 algorithm,and its detection speed reaches 48 fps and 45 fps respectively,which ensure the real-time performance and greatly improve the detection accuracy.

关 键 词：遥感图像 目标检测 跨阶段残差结构 特征交互 跨阶段分层卷积模块 激活函数 非线性特征 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]