基于特征融合SSD的遥感图像舰船目标检测

Ship target detection in remote sensing images based on feature fusion SSD

针对当前常用的舰船目标检测算法的精度与实时性兼顾不佳的问题,提出了基于特征融合的目标检测算法FFSSD,该算法以SSD算法和DSSD为基础,克服了SSD算法小目标检测精度低、DSSD算法运行时间长等缺陷。FFSSD算法选取轻量级VGG16网络作为特征提取的骨干网络,然后设计特征融合模块,将深层特征通过反卷积操作与浅层特征进行融合,得到结构层次丰富的融合特征,最终实现不同卷积层、不同特征的多元信息的分类预测与位置回归。在舰船光学遥感数据集上进行数据分析实验,平均检测精度(AP)达到85.9%,检测速度达到38.6FPS。实验结果表明,FFSSD算法能在实现实时检测的同时有效提高检测精度,实现对遥感图像中舰船目标的高精度检测。

Concerning the poor accuracy and real-time performance of commonly used ship target detection methods,a feature fusion target detection algorithm named FFSSD(Feature Fusion Single Shot Detector)was proposed.Based on SSD(Single Shot multibox Detector)and DSSD(Deconvolutional-SSD),FFSSD overcomes the inaccurate recognition in small target of SSD and long running time of DSSD.In FFSSD,the lightweight VGG16 network was selected as the backbone network for feature extraction.Then the feature fusion module was designed to fuse the deep features and the shallow features by using deconvolution operation to obtain the fusion features with rich structure levels.Finally,classification prediction and position regression of multi-information with different convolutional layers were realized.Data analysis experiment was carried out on a ship optical remote sensing dataset,the Average Precision(AP)of FFSSD reached 85.9%,and the detection speed reached 38.6 FPS(frame/second).Experimental results show that FFSSD can effectively improve the detection accuracy with satisfying real-time detection,and achieve high-precision detection of ship targets in remote sensing images.