基于特征增强的光学遥感图像舰船目标检测算法

Ship target detection algorithm based on feature enhancement in remote sensing

针对光学遥感图像背景复杂、特征提取难度高导致舰船检测精度低的问题,基于包围框边缘感知向量(box boundary-aware vectors,BBAVectors)网络,提出了一种特征增强的遥感舰船检测算法。为适应舰船的多尺度特性,利用最大池化和空洞卷积构建了增强型感受野模块(enhanced receptive file block,ERFB),增加模型感受野,同时提高对小尺度目标的特征提取能力;为提高各层特征的利用率,设计了一个自适应特征融合模块(adaptive feature fusion module,AFFM),在丰富特征的同时,突出相对重要的特征。在公开数据集HRSC2016和自建数据集RS-Ship上进行实验,与原始BBAVectors相比,所提方法的AP值分别提高了5.68%和7.84%,有效提高了舰船检测的准确率;与其他方法相比,所提方法在精度上表现优异,并拥有较强的鲁棒性。

To address the problem of low ship detection accuracy due to complex background and difficult feature extraction of optical remote sensing images,a feature enhanced ship detection algorithm based on box boundary-aware vectors(BBAVectors)network is proposed in this paper.In order to suit the multi-scale characteristics of ship,an en-hanced receptive file block(ERFB)is constructed by using max pooling and dilated convolution to increase the recep-tive field of the model and improve the feature extraction capability for small-scale targets.To improve the utilization of features in each layer,an adaptive feature fusion module(AFFM)is designed to enrich the features while highlighting the relatively important features.Experiments are conducted on the public dataset HRSC2016 and the self-built dataset RS-Ship,and the AP values of the proposed method are improved by 5.68%and 7.84%,respectively,compared with the original BBAVectors,which effectively improves the accuracy of ship detection.Compared with other methods,the proposed method performs well in terms of accuracy and has strong robustness.