基于改进YOLOV3算法的遥感图像油罐检测

Oil tank detection in remote sensing images based on improved YOLOV3 algorithm

针对遥感图像中的油罐检测问题,借鉴深度神经网络的思想,提出了一种基于改进YOLOV3算法的图像检测方法。首先在原有YOLOV3算法框架中添加空间变换网络(spatial transformer networks,STN),使其成为具备空间变换学习能力的模型;然后通过k-means聚类算法对数据集进行分析,重新设计初始候选框大小;最后训练和测试网络,建立包含9724个油罐目标的遥感图像数据集。实验结果表明:改进的YOLOV3算法具有良好的性能,在测试集中的召回率可达到95.64%,比原算法提升了3.52%;准确率可达到93.92%,比原算法提升了2.81%。

In order to solve the problem of oil tank detection in remote sensing images,an improved image detection method based on YOLOV3 algorithm was proposed by referring to the idea of deep neural network.Firstly,a spatial transformer network(STN)was added to original YOLOV3 framework to make it a model with spatial transformation learning ability.Then,k-means clustering algorithm was used to analyze the dataset and redesign the size of the initial candidate box.Finally,a remote sensing images dataset including 9724 targets of oil tank was built to train and test the CNNs.The experimental results show that the improved YOLOV3 algorithm has good performance.The recall rate in the test set reaches 95.64%,which is 3.52%higher than that of the original algorithm;The accuracy rate reaches 93.92%,which is 2.81%higher than that of the original algorithm.