摘要
为了从光学遥感卫星图像中快速提取船舶目标显著性区域,同时为解决显著性数据集稀缺以及标注困难的问题,文中设计了一种多尺度全卷积神经网络,实现了基于眼动注视数据的船舶目标显著性图的生成,并针对遥感图像舰船检测任务,利用Google Earth采集包含舰船的高分辨率光学遥感图像,再使用TobiiProX3-120台式眼动仪构建舰船目标显著性检测数据集。实际眼动数据生成的真值图经过训练测试后发现,文中所提方法的陆地移动距离系数(EMD)达到0.003,从定性评价和定量评价上看,明显优于深度监督模型(DVA)等显著性检测模型。
In order to quickly extract the salient region of ship target from optical remote sensing satellite image and at the same time,in order to solve the problem of the scarcity of significant data sets and the difficulty of annotation,we design a multi-scale full convolution neural network to generate the ship target saliency map based on the eye movement fixation data.For the ship detection task of remote sensing image,we use Google Earth to collect the high-resolution optical remote sensing image containing the ship,and then use TobiiProX3-120 desktop eye tracker to build the ship target saliency detection data set.After training and testing,it is found that the proposed method achieves 0.003 in earth move distance(EMD).From the qualitative and quantitative evaluation,it is obviously better than the other significance detection models such as Deep Visual Attention prediction model(DVA).
作者
曾禹龙
ZENG Yulong(Shanghai Jiao Tong University,Shanghai 200240,China)
出处
《电子设计工程》
2022年第8期52-56,共5页
Electronic Design Engineering
关键词
遥感图像
显著性检测
眼动数据
CNNS
remote sensing image
saliency detection
eye movement data
CNNs
