融合多层特征与上下文信息的YOLO改进算法  

Improved YOLO Algorithm via Fusing Multilayer Features and Contextual Information

作　　者：费选[1] 郭梦瑶 吴思佳 靳子泷 马丁[1] FEI Xuan;GUO Meng-yao;WU Si-jia;JIN Zi-long;MA Ding(School of Artificial Intelligence and Big Data,Henan University of Technology,Zhengzhou 450001,China)

机构地区：[1]河南工业大学人工智能与大数据学院,郑州450001

出　　处：《科学技术与工程》2025年第4期1555-1562,共8页Science Technology and Engineering

基　　金：国家自然科学基金青年科学基金(62006072);河南省重点研发与推广专项(科技攻关)项目(222102210108);粮食处理与控制教育部重点实验室开放课题(KFJJ2022013);河南工业大学创新基金支持计划专项资助(2022ZKCJ11);河南工业大学青年骨干教师培育计划。

摘　　要：遥感图像目标检测在军事侦察、智慧农业等领域意义重大,特别是小目标检测一直获得持续关注。然而,遥感图像中的小目标面临特征信息不足、检测难度大等问题,成为困扰遥感检测应用发展的最大障碍。为此,提出YOLO-HF(you only look once-hybrid feature)算法,该算法在传统YOLOv7模型的网络中,引入通道注意力和自注意力的混合注意力机制提取目标深层特征,并将浅层特征和深层特征进行融合,增加局部特征的丰富性;为进一步加强对全局信息的关注,在提取特征后为小尺度目标添加全局注意力机制,实现全局特征表达能力的提升;为避免传统损失函数对小目标位置偏差敏感,导致检测效果不佳,选择使用一种新的度量方式,将其嵌入边界框损失函数的计算中,从而加快损失函数的收敛,实现小目标检测精度的提升。实验结果表明:与传统YOLOv7算法相比,所提算法在RSOD和NWPU VHR-10数据集上均表现出优越性,特别地,在RSOD数据集上均值平均精度提升了2.90%,在NWPU VHR-10数据集上均值平均精度实现了3.61%的提升。Remote sensing image target detection is one of great significance in military reconnaissance,intelligent agriculture and other fields,especially small target detection has been gaining continuous attention.However,small targets in remote sensing images face the problems of insufficient feature information and difficult detection,which have become the biggest obstacles plaguing the development of remote sensing applications.To this end,the you only look once-hybrid feature(YOLO-HF)algorithm was proposed,which introduced a hybrid attention mechanism of channel attention and self-attention in the network of the traditional YOLOv7 model to extract the target s deep features,and fused the shallow and deep features to increase the richness of local features;to further strengthen the attention to the global information,a global attention mechanism was added for the small-scale targets after the extraction of the features,to achieve the ability of global feature expression enhancement.In order to avoid that the traditional loss function was sensitive to the positional deviation of small targets,which leaded to poor detection effect,a new metric was selected for use,which was embedded into the computation of the bounding box loss function,so as to accelerated the convergence of the loss function and realized the enhancement of the detection accuracy of small targets.The experimental results show that compared with the traditional YOLOv7 algorithm,the proposed algorithm shows superiority on both RSOD and NWPU VHR-10 datasets,and in particular,the mean average accuracy on RSOD dataset is improved by 2.90%,and the mean average accuracy on NWPU VHR-10 dataset realizes an improvement of 3.61%.

关 键 词：遥感图像 目标检测 YOLOv7 多层特征 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]