Research On Neural Network For RGB-D Saliency Object Detection

Saliency Object Detection(SOD)aims to display visually salient areas in an image.Thanks to this ability to capture regions of human interest,saliency object detection plays an important role in many computer vision tasks.Representative tasks and applications include image understanding,semantic segmentation,visual object tracking,and video compression,etc.The wide application of depth sensors has promoted the development of RGB-D saliency object detection.Most existing RGB-D SOD methods mainly use a symmetric two-stream architecture.However,due to inherent differences between RGB data and depth data,it is suboptimal to extract RGB and depth features using a symmetric architecture.How to design a architecture fit better to RGB-D saliency object detection,how to effectively extract rich global context information while preserving local saliency details,and how to effectively utilize the discriminative power of depth features to guide the RGB features for locating salient objects accurately has become the key issues to be addressed in the saliency object detection field.Aiming at the above problems,this paper proposes an asymmetric two-stream architecture.Specifically,the Flow Ladder Module(FLM)designed in the RGB branch not only retains saliency details but also comprehensively extracts global and local information.The FLM consists of four detail transfer branches,each of which receives global position information from the representation of other vertical parallel branches in an evolutionary way while retaining the detail information.In addition,this paper proposes a novel Depth Attention Module(DAM),which selectively utilizes depth information to guide RGB features to ensure the effective role of RGB information in challenging scenes.This paper proposed model is evaluated with thirteen state-of-the-art methods on six benchmark RGB-D saliency detection datasets.A large number of experiments show that compared with other methods,this method achieves more accurate and high-quality results.Among the five evaluation metrics,the performance of the method in this paper is consistently better than the other 13 latest RGB-D methods.