多尺度特征增强的遥感图像舰船目标检测

Ship detection based on multi-scale feature enhancement of remote sensing images

针对背景复杂的遥感图像中,舰船方向任意、密集排列造成的漏检问题,基于旋转区域检测网络,提出多尺度特征增强的遥感图像舰船目标检测算法。在特征提取阶段,利用密集连接感受野模块改进特征金字塔网络,选用不同空洞率的卷积获取多尺度感受野特征,增强高层语义信息的表达;为了抑制噪声并突出目标特征,在特征提取后设计基于注意力机制的特征融合结构,根据各层在空间上的权重值融合所有层,得到兼顾语义信息和位置信息的特征层,再对该层特征进行注意力增强,将增强后的特征融入原金字塔特征层;在分类和回归损失基础上,增加注意力损失,优化注意力网络,给予目标位置更多关注。在DOTA遥感数据集上的实验结果表明,该算法平均检测精度可以达到71.61%,优于最新的遥感图像舰船目标检测算法,有效地解决了目标漏检问题。

Aiming at the omission in the ship target detection from remote sensing images with complex background caused by the arbitrary and dense arrangement of ships,this study,based on the rotation region generation network,proposes a ship target detection algorithm using the multi-scale feature enhancement of remote sensing images.The detailed steps are as follows.Firstly,improve the feature pyramid network using the receptive field module with dense connection at the feature extraction stage.Then obtain the characteristics of multi-scale receptive fields using the convolution of different dilate rates.In this way,the expression of high-level semantic information can be enhanced.Then design a feature fusion structure based on attention mechanisms to restrain noise and highlight the target characteristics.Afterward,fuse all layers according to the spatial weight value of each layer to obtain a feature layer that takes both semantic and position information into account.Then conduct attention enhancement to the features of this layer,and integrate the enhanced features into the original feature layer in the pyramid network.Consequently,pay more attention to target locations by increasing attention loss and optimizing the attention network according to the classification and regression loss.As indicated by the experiment results of DOTA remote sensing dataset,the average precision of this algorithm is as high as 71.61%,which is higher than the latest ship target detection algorithm based on remote sensing images.In this manner,the omission in ship target detection can be effectively solved.