基于改进YOLOv5的光学遥感图像水坝检测研究

Research on Dam Detection of Optical Remote Sensing Image Based on Improved YOLOv5

目标检测是计算机视觉领域的一个重要应用,针对光学遥感影像的目标检测任务也是当下的研究热点之一。现阶段科技进步的同时带来了一系列环境问题,环境保护已经成为当下值得关注的重点问题。水坝的建设是影响全球环境保护以及资源利用的一个重要因素,对水坝进行监测可以为环境保护工作提供参考依据。为了环境保护后续工作的开展,分析水坝在图像中的位置,该文针对高分辨率光学遥感影像中的水坝目标检测方法进行研究,对比了深度学习三个阶段较为典型的目标检测模型,根据实验结果选用精度较高的YOLOv5通用目标检测模型,并根据遥感图像背景复杂的特性结合CBAM注意力机制提高网络对图像中水坝目标的重点关注。在DIOR光学遥感目标检测数据集中提取含有水坝目标的图像并验证模型精度,实验表明YOLOv5-CBAM在并不显著增加模型大小的情况下比YOLOv5运算能力强,并且AP50可以达到86.4%,比仅使用YOLOv5的模型AP50提高了3.2百分点。

Object detection is an important application in the field of computer vision,and the task of object detection for optical remote sensing images is also one of the current research hotspots.At this stage,the progress of science and technology has brought a series of environmental problems at the same time,so environmental protection has gradually become a key issue worthy of attention.The construction of dams has become an important factor affecting global environmental protection and resource utilization.Monitoring of dams can provide reference for environmental protection work.In order to carry out the follow-up work of environmental protection and analyze the position of the dam in the image,we study the dam target detection method in high-resolution optical remote sensing images,and compare the typical target detection models in three stages of deep learning.According to the experimental results the YOLOv5 general target detection model with higher accuracy is selected,and according to the complex characteristics of the remote sensing image background combined with the CBAM attention mechanism,the network’s focus on the dam target in the image is improved.Extract images containing dam targets in the DIOR optical remote sensing target detection dataset and verify the model accuracy.Experiments show that YOLOv5-CBAM has stronger computing power than YOLOv5 without significantly increasing the model size,and the AP50 of YOLOv5 can reach 86.4%,which is 3.2 percentage points higher than that of the model using only YOLOv5.