特征增强SSD算法及其在遥感目标检测中的应用

Feature Enhancement SSD Algorithm and Its Application in Remote Sensing Images Target Detection

为了提高对复杂场景下多尺度遥感目标的检测精度,提出了基于多尺度单发射击检测(SSD)的特征增强目标检测算法.首先对SSD的金字塔特征层中的浅层网络设计浅层特征增强模块,以提高浅层网络对小目标物体的特征提取能力;然后设计深层特征融合模块,替换SSD金字塔特征层中的深层网络,提高深层网络的特征提取能力;最后将提取的图像特征与不同纵横比的候选框进行匹配以执行不同尺度遥感图像目标检测与定位.在光学遥感图像数据集上的实验结果表明,该算法能够适应不同背景下的遥感目标检测,有效地提高了复杂场景下的遥感目标的检测精度.此外,在拓展实验中,文中算法对图像中的模糊目标的检测效果也优于SSD.

In order to improve the detection accuracy of multi-scale remote sensing ship targets in complex scenes,a feature enhancement single shot multi-scale detector is proposed.Firstly,the shallow feature enhancement module is designed to improve the feature extraction ability of the shallow network in the pyramid structure of Single Shot MultiBox Detector(SSD).Then the deep feature fusion module is designed to replace the deep network in the pyramid structure of SSD to improve the feature extraction ability of deep network.Finally,the image features are matched with candidate frames of different aspect ratios to adapt to remote sensing image targets of different scales.The experiments tested on the optical remote sensing image dataset demonstrate that the proposed method can adapt to target detection under different background and effectively improve the detection performance of multi-scale remote sensing targets in complex scenes.On the extended experiment,the proposed method performance over SSD in blurry target detection.