Research And Implementation Of RGBD Salient Object Detection Algorithm

Salient object detection aims to simulate human to locate and segment the most visually distinctive objects or regions in a given image.This task serves as an important branch in computer vision community,and it is widely used in many actual scenes,such as robot recognition,background conversion,and 3D vision reconstruction.Moreover,it is also as a fundamental step in various vision tasks,including image recognition,image classification,and semantic segmentation.In recent years,with the rapid development of deep learning,CNNs based approaches have achieved remarkable performance in RGBD salient object detection.But how to effectively exploit depth clues is still a challenge.Firstly,depth images captured from acquisition devices(such as Kinect,i Phone X)often contain noises,which makes the extraction of depth features challenge.Secondly,existing deep learning methods directly integrate RGB and depth cues.However,due to the inherent inconsistent between RGB and depth information,the RGB features are easy to be interfered by the intrinsic noise existed in depth features,making the precise RGBD saliency detection still a challenge.This paper focuses on solving the above two issues and the main contributions are summarized as follows:(1)RGBD Salient Object Detection based on Adaptive Weight.Aiming at solving the problem of two different attribute features fusion,this paper proposes a novel dynamic depth fusion network(D2FNet)that takes depth information as a prior and dynamic selects the complementary RGB information for RGBD salient object detection.Specifically,D2 FNet first learns a set of layer-specific weights from multi-scale depth features.Guiding by these learned weights,D2 FNet further assigns them to their corresponding RGB layers for dynamically enhancing and selecting saliency-related RGB features.With these two steps,D2 FNet is able to effectively integrate the multimodality complementaries and further highlight salient regions.(2)RGBD Salient Object Detection based on Edge Information.Aiming at solving the issue of feature extraction challenge that caused by the presence of noise data in depth images(mainly manifested as edge blur),this paper proposes a novel edge learning based network(BELNet)for RGBD salient object detection.Concretely,BELNet designs a depth-edge extraction module to acquire the edge information of depth image.It then devises a category-balanced cross-entropy loss function to ensure the accuracy of depth edge information extraction.With this,BELNet solves the problem of inaccurate depth feature extraction caused by the blur of depth image edges to a certain extent,and it can segment salient objects with higher accuracy.In order to verify the effectiveness of the proposed methods,this paper constructs comprehensive experiments on several publicly used RGBD salient object detection datasets.Experimental results demonstrate that our two proposed methods can fairly locate salient regions and effectively segment the complete object.