Font Size: a A A

Research On RGB-D Salient Object Detection With Low-Quality Depth Map

Posted on:2023-04-20Degree:MasterType:Thesis
Country:ChinaCandidate:X ZhengFull Text:PDF
GTID:2568307022997809Subject:Software engineering
Abstract/Summary:PDF Full Text Request
Salient Object Detection(SOD)recognizes the most distinct objects in the scene by imitating human visual perception,which is a basic task in the field of computer vision.However,the traditional SOD method is based only on RGB images.When in a complex scene,these methods may obtain incomplete detection results,confusion of targets and backgrounds,or misjudgment of salient regions due to the blurred borders and dark brightness.The depth map contains geometric information,With the appearance of the depth map,it can provide extra spatial information for the RGB-based SOD,making the final saliency map more accurate.However,the introduction of depth maps will trigger two problems,one is the low quality of the depth map,and the other is the difficulty of the cross-modal information fusion problem.The low quality of depth map will bring interference information to the neural network,which will reduce the detection accuracy.Moreover,there is a semantic gap between the color map and the depth map,which will make it hard to fuse the cross-modal feature,and when the quality of the depth map is low,the salient features learned through the depth map are unreliable,thus the predicted result will be more unreasonable.As for the depth map with low-quality may bring interference information into the network,histogram intersection method is proposed to obtain the confidence score of each depth map.This method can not only quantitatively evaluate the quality of the depth map,but also solve the inconsistency of depth format to some extent.The confidence score of the depth map is used to control the participation of the low-quality depth map,thus reducing bringing the interference information.As for the difficulty of cross-modal information fusion problem,Cross Modal Fusion Block(CMFB)module is proposed to reduce the semantic gap,which enhances the learning of salient features through mutual attention.This mutual attention mechanism can help get better fusion results between RGB and depth modal.Further,an architecture is designed for RGB-D salient object detection with the low quality depth maps based on the confidence score and the CMFB module.The experimental results have shown great performance on 6 datasets,especially on the DES,compared with the DRLF methods,Sα、Fβ and Eγ increase by 4.2%,3.7%,and 5.8%respectively,which has shown the effectiveness of the module and the network.
Keywords/Search Tags:RGB-D Salient Object Detection, Depth Map, Confidence Score, Attention Mechanism, CNN
PDF Full Text Request
Related items