摘要
YOLOv7是目前目标检测任务中性能较优的模型,但在处理遥感影像中的道路交叉口时,出现目标背景复杂、先验框定位误差以及模型训练参数量增多的问题。针对复杂场景的道路交叉口提出一种结合归一化高斯Wasserstein距离与轻量级YOLOv7的遥感影像道路交叉口检测模型。首先,使用归一化高斯Wasserstein距离与CIoU(complete-IoU)进行先验框定位损失函数的改进,以提高网络模型对于目标尺寸的鲁棒性;其次,在加强网络特征提取模块中加入三维注意力机制,实现网络处理的特征优化;最后,在主干特征提取网络与加强特征提取网络中加入改进的FasterNet模块,提升网络模型的训练速度,减少了模型训练的参数。实验结果表明,改进后的YOLOv7网络模型相比原网络模型,漏检测情况得到明显改善,准确率(precision,P)、召回率(recall,R)、平均准确率(average precision,AP)和F_(1)分别提升了6.2%、4.9%、6.7%、6.5%,对道路交叉口的检测效果优于原网络模型。其成果对不同环境的影像具有较强适应能力,为道路交叉口检测的发展提供了参考。
YOLOv7 is a better-performing model in the current target detection task.However,when dealing with road intersections in remote sensing images,the problems of complex target backgrounds,significant positioning errors in the first frame,and the number of model training parameters increase.A road intersection detection model for remote sensing images that combines normalized Gaussian Wasserstein distance with lightweight YOLOv7 was proposed for road intersections in complex scenarios.Firstly,the transcendental box localization loss function was improved by normalized Gaussian Wasserstein distance and complete-IoU(CIoU)to enhance the robustness of the network model to target size.Secondly,a three-dimensional attention mechanism was added to the enhanced network feature extraction module to achieve feature optimization.Finally,an improved FasterNet module was added to the backbone feature extraction network and enhanced feature extraction network to improve the network model s training speed and reduce the model training parameters.The results show that the improved YOLOv7 network model shows significant improvement in leakage detection compared to the original network model,with precision(P),recall(R),average precision(AP),and F_(1)values increasing by 6.2%,4.9%,6.7%,and 6.5%,respectively.It is concluded that solid adaptability to different environmental images provides a reference for developing road intersection detection.
作者
康传利
张思瑶
李玄皓
林梓涛
耿崇铭
张赛
王世伟
KANG Chuan-li;ZHANG Si-yao;LI Xuan-hao;LIN Zi-tao;GENG Chong-ming;ZHANG Sai;WANG Shi-wei(College Geomatics and Geoinformation,Guilin University of Technology,Guilin 541006,China)
出处
《科学技术与工程》
北大核心
2024年第9期3533-3542,共10页
Science Technology and Engineering
基金
国家自然科学基金(41961063,42064002)。
