改进YOLOv3算法及其在航拍图像车辆检测中的应用  

IMPROVED YOLOV3 ALGORITHM AND ITS APPLICATION ON AERIAL IMAGE VEHICLE OBJECT DETECTION

作　　者：丛眸 张平 王宁[3] Cong Mou;Zhang Ping;Wang Ning(Institute of Computer Science and Technology,Changchun University of Science and Technology,Changchun 130022,Jilin,China;Army Academy of Armored Forces,Beijing 100072,China;The Third Research Institute of Ministry of Public Security,Beijing 100142,China)

机构地区：[1]长春理工大学计算机科学技术学院,吉林长春130022 [2]陆军装甲兵学院,北京100072 [3]公安部第三研究所,北京100142

出　　处：《计算机应用与软件》2023年第1期228-233,共6页Computer Applications and Software

摘　　要：针对航拍图像中的车辆目标尺度小、特征不明显导致目标检测困难的问题,提出一种改进YOLOv3的航拍车辆目标检测方法。将空间金字塔池化模块引入到特征提取网络中,丰富卷积特征的表达能力;设计4个不同尺度的卷积特征金字塔,并通过卷积特征融合机制来实现对多层级卷积特征的融合,在融合后的卷积特征金字塔上进行目标检测。在航拍图像车辆目标检测数据集上的测试结果表明,与原YOLOv3相比,改进后的算法能够有效地提高对航拍图像中车辆目标检测效果的查全率以及查准率,并将平均均值精度(mean average precision, mAP)提升了4.5百分点。The target scale of vehicles in aerial images is small and the features are not obvious, which make it difficult to detect targets. Aiming at these problems, we propose an improved YOLOv3 aerial images vehicle object detection method. The spatial pyramid pooling module was introduced into the feature extraction network to enrich the expression ability of convolution features. We designed a convolution feature pyramids with four different scales, and realized the fusion of multi-level convolution features through the convolution feature fusion mechanism. The object detection was performed on the fused convolution feature pyramid. The test results on the aerial image vehicle object detection data set show that compared with the original YOLOv3, the improved algorithm can effectively improve the recall rate and precision rate of the vehicle target in the aerial image, and improve the average mean precision by 4.5 percentage points.

关 键 词：车辆检测 航拍图像 YOLOv3 空间金字塔池化 卷积特征融合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]