基于非对称跨模态融合的RGB-D显著目标检测

RGB-D salient object detection with asymmetric cross-modal fusion

目前大多数RGB-D显著目标检测方法在RGB特征和Depth特征的融合过程中采用对称结构,对两种特征进行相同的操作,忽视了RGB图像和Depth图像的差异性,易造成错误的检测结果.针对该问题,提出一种基于非对称结构的跨模态融合RGB-D显著目标检测方法,利用全局感知模块提取RGB图像的全局特征,并设计了深度去噪模块滤除低质量Depth图像中的大量噪声;再通过所提出的非对称融合模块,充分利用两种特征间的差异性,使用Depth特征定位显著目标,用于指导RGB特征融合,补足显著目标的细节信息,利用两种特征各自的优势形成互补.通过在4个公开的RGB-D显著目标检测数据集上进行大量实验,验证所提出的方法优于当前的主流方法.

Most RGB-D salient object detection methods use a symmetric structure during the fusion process to perform the same operation on the RGB features and Depth features.This fusion method ignores the difference between the RGB image and the Depth image,which is likely to cause false detection results.In order to solve it,this paper proposes a cross-modal fusion RGB-D salient object detection method based on an asymmetric structure.In this paper,a global perception module(GPM)is designed to extract the global features of RGB images,and a deep denoising module(DDM)is designed to filter out a large amount of noise in low-quality depth images.Then through the asymmetric fusion module designed,we make full use of the difference between the two features differences,use the depth feature to locate salient objects,so as to guide RGB feature fusion and complement the detailed information of salient objects,and use the respective advantages of the two features to form a complement.A large number of experiments are carried out on four publicly available RGB-D salient object detection datasets,and the experimental results verify that the proposed method outperforms the state-of-the-art methods.