Algorithms Of RGB-D Image Salient Object Detection

As known,computer vision plays a very important role in the development of Computer Science in recent years because of its special position in replacing human eye recognition with cameras and other devices.As a preprocessing step of computer vision,salient object detection has very high research value.The task of salient object detection is to extract the part of the image that the human eye may be interested in by the algorithm.Essentially,it is to judge that each pixel in the image belongs to the classification of foreground or background.Salient object detection plays an important role in astronomy,medicine,finishing,public safety and military fields.At present,salient object detection can be divided into four fields:RGB co-saliency detection,RGBD co-saliency detection,RGB single image salient object detection and RGBD single image salient object detection.The research direction of this paper is RGBD single image salient target detection.RGBD image salient target detection aims to use the information provided by a pair of RGB images and depth images to calculate the most visually salient target.In recent years,machine learning and its derivative Deep Learning have been widely used in the field of salient target detection.The key point is to train a classification model by using label training samples.Unfortunately,many excellent and potential classifiers are only widely used in the field of RGB salient target detection,but they have not been fully developed in the field of RGBD salient target detection.The computational flow of some machine learning algorithms and the correction of salient maps also have perfect space.Based on these problems,this paper has done two aspects of work.Firstly,this paper explores Support Vector Machine(SVM)and Extreme Learning Machine(ELM).The RGB image is put into the trained feature extraction model,and the high-dimensional depth feature information(4096 dimensions)is extracted by CNN.At the same time,several low-level features,such as RGB,Depth,LBE,are extracted to describe the color,texture,depth and shape features of image blocks as eigenvalues together with high-dimensional information features.This is RGBD salient object detection via ELM.Secondly,this paper explores the nature of depth information,mainly analyzes the problem of which situation depth map may fail,and studies the application of depth map hierarchy in specific scenarios.This is Depth image salient object detection based on vanishing point,and combining with ELM framework,the depth feature is extended to RGBD salient object detection.In this paper,RGBD image processing is tested on RGBD-1000 and NJUDS-2000 dataset.The test results show that it has more advantages than the most advanced salient object detection algorithms.