采用改进YOLOv3的高分辨率遥感图像目标检测

由于高分辨率遥感图像存在目标排列密集、尺寸差别大等情况，传统算法难以准确地对其进行目标检测。在YOLOv3算法的基础上，提出一种改进的高分辨率遥感图像目标检测算法(remote sensing-YOLO，RS-YOLO)。利用K-means聚类算法对数据集进行聚类，重新设计适合遥感图像的先验框; 引入高斯模型计算预测框的不确定度，以提高网络对预测框坐标的准确度; 使用弱化的非极大值抑制算法(soft non-aximum suppression，Soft-NMS)对预测框进行处理，增强算法对密集排列目标的检测能力。实验结果表明，改进后的算法能够对高分辨率遥感图像进行有效的目标检测，以NWPU VHR-10数据集为例，RS-YOLO的平均检测精度达到了87.97%。

Object detection of high resolution remote sensing images based on improved YOLOv3

Due to the dense arrangement of targets and the big difference of target sizes, it is difficult to detect objectives accurately in high-resolution remote sensing images by traditional algorithms. An improved high-resolution remote sensing image object detection method (RS-YOLO) based on YOLOv3 algorithm is proposed. K-means algorithm is used to cluster data set and redesign the prior anchor box suitable for remote sensing image. Gaussian model is introduced to calculate the uncertainty information of bounding box to improve the accuracy of coordinates from the network. Soft-NMS algorithm is used to process the bounding box which enhance the detection ability for densely arranged targets. Experimental results show that the improved algorithm is effective in high-resolution remote sensing images object detection. In NWPU VHR-10 data set, the average detection precision of RS-YOLO reached 87.97%.