摘要
伪装目标分割的任务是使用像素级分割掩码将与背景高度相似的目标进行准确分类和定位,与传统的目标分割任务相比更具挑战性。针对目标与周围环境高度相似、边界模糊、对比度低等问题,构建了一种基于边缘增强和特征融合的伪装目标分割方法。首先,设计了一组边缘提取模块,能够更准确地分割有效的边缘先验。之后,引入了多尺度特征增强模块和跨层级特征聚合模块,分别挖掘层内与层间的多尺度上下文信息。提出了一种简单的层间注意力模块,利用相邻层级间的差异有效滤除融合后存在的干扰信息。最后,通过将各级特征图与边缘先验逐级结合的方式,获得准确的预测结果。实验结果表明,在4个伪装目标基准数据集上,该模型的表现都优于其他算法。其中加权F值提升了2.4%,平均绝对误差减少了7.2%,在RTX 2080Ti硬件环境下分割速度达到了44.2 FPS。与现有方法相比,该算法能够更准确地分割伪装目标。
The task of camouflaged object segmentation is to accurately classify and localize objects that are highly similar to the background using pixel-level segmentation masks,which is more challenging than traditional object segmentation tasks.Aiming at the problems that the target is highly similar to the surrounding environment,the boundary is blurred,and the contrast is low,a camouflaged target segmentation method based on edge enhancement and feature fusion is constructed.First,a set of edge extraction modules is designed,aiming to accurately segment valid edge priors.Afterwards,a multi-scale feature enhancement module and a cross-level feature aggregation module are introduced to mine multi-scale contextual information within and between layers,respectively.In addition,a simple inter-layer attention module is proposed to effectively filter out the interference information existing after fusion by utilizing the difference between adjacent layers.Finally,accurate prediction results are obtained by combining feature maps of all levels with edge priors step by step.Experimental results show that the model outperforms other algorithms on four camouflaged target benchmark datasets.Among them,the weighted F value increased by 2.4%,the average absolute error decreased by 7.2%,and the segmentation speed reached 44.2 FPS under the RTX 2080Ti hardware environment.Compared with existing methods,this algorithm can segment camouflage targets more accurately.
作者
李明岩
吴川
朱明
LI Mingyan;WU Chuan;ZHU Ming(Changchun Institute of Optics,Fine Mechanics and Physics,Chinese Academy of Sciences,Changchun 130033,China;University of Chinese Academy of Sciences,Beijing 100049,China)
出处
《液晶与显示》
CSCD
北大核心
2024年第1期48-58,共11页
Chinese Journal of Liquid Crystals and Displays
关键词
深度学习
伪装目标
图像分割
边缘特征
特征融合
deep learning
camouflaged object
image segmentation
edge feature
feature fusion
