基于全卷积神经网络的遥感图像海面目标检测

Sea Surface Target Detection in Remote Sensing Images Based on Full Convolution Neural Network

针对通常的神经网络算法在检测遥感图像海面目标时存在精确率低、漏检概率高的问题,改进了一种基于YOLOv3全卷积神经网络的遥感图像海面目标检测方法。首先根据海面目标的宽高比例,利用Kmeans++聚类算法,确定出适合于数据集的anchor box值;接着采用FPN思想进行特征融合;最后,选用GIOU作为坐标预测的损失函数,进一步优化检测结果。实验表明:文中方法在遥感图像海面目标检测中的平均精确率为90.82%,相比于其他算法平均提高了5.34%。

In order to solve the problems of low accuracy and high probability of missing detection in the detection of remote sensing image sea surface targets,a new method of remote sensing image sea surface targets detection based on YOLOv3 full convolution neural network is improved.Firstly,according to the ratio of width to height of sea surface targets,the anchor box value suitable for data sets is determined by Kmeans++clustering algorithm;then,the feature fusion is carried out by using FPN idea;finally,GIOU is selected as the loss function of coordinate prediction to further optimize the detection results.Experiments show that the mean Average Precision(mAP)of this method is 90.82%in the sea surface targets detection of remote sensing images,which is 5.34%higher than other algorithms.