Research On Optimization Method Of RGB-D Saliency Map

Computer vision is a kind of simulation engineering of biological vision using computer or related equipment.It is a scientific subject that studies the machine's ability to identify,detect,and track objects in images instead of the human eyes.As one of the branches of computer vision,visual saliency detection is to simulate human visual characteristics and extract salient regions from images through intelligent algorithms.RGB-D saliency detection is designed to identify the most visually significant objects in a pair of RGB images and depth images.In the past few decades,a number of scholars or teams have researched and published a variety of RGB-D image saliency detection algorithms in the academic world.These research have achieved remarkable results and promoted rapid progress in this field,but there are still part of the algorithm,whose detection accuracy or efficiency is still lacking.Therefore,based on the existing RGB-D saliency maps,considering the case that the number of initial saliency maps is single or multiple,two optimization methods for RGB-D salient images in two different situations is proposed.In recent years,machine learning has been widely applied to the field of image saliency detection.The key point is to train a classification model by training samples with manually labeled tags.A large number of training samples directly affect the performance of the learning model.RBF(radial Basis Function)network could approximate arbitrary nonlinear functions,could deal with difficult analytical regularity in the system,it has good generalization ability and fast learning convergence speed.It has been successfully applied for nonlinear function approximation,data classification,pattern recognition,information processing,image processing and so on.Therefore,an effective RGB-D saliency map optimization algorithm based on RBF model is proposed in this paper.In addition,the academic community has proposed various effective methods for single RGB-D image saliency detection,and these detection methods have complementary advantages.Therefore,the studies that the fusion of the saliency maps generated by various methods to improve the accuracy of saliency detection are indispensable.As the name implies,the saliency map fusion work fuses the saliency maps generated by various image saliency detection methods to obtain the final saliency map.Many scholars have done a lot of research work in this field,and could prove that the fusion work can achieve the significant results.Therefore,for the case where the initial saliency map is multiple,this paper proposes a fusion algorithm of RGB-D saliency map.The main research work of this paper has the following two points:1.This paper presents a RGB-D saliency map optimization algorithm based on RBF model.Firstly,the saliency map generated by the existing RGB-D image saliency detection algorithm is regarded as a weak saliency map,that is,a priori map,and a training set containing positive and negative samples is collected from the superpixels using the priori map.Secondly,four feature descriptors RGB,CIElab,LBP(Local Binary Pattern)and depth value DEPTH are extracted from the input RGB image and depth image to perform the rich feature representation,so as to learn a strong classifier to map the eigenvectors and saliency values of the superpixels,and then uses the trained RBF model to predict the saliency of the testing samples to obtain the labeled saliency map.Finally,in order to further improve the detection performance,the multi-scale labeled saliency map are combined to generate a strong saliency map,and then the weak saliency map and the strong saliency map are combined to obtain the final RGB-D image saliency detection result.2.In this paper,the RGB-D image saliency map is fused,that is,in the process of merging the saliency map,the depth map is used for influence.in this fusion algorithm.Saliency maps are fused at the image level and the pixel level,respectively.At the image level,a simple and effective feature of RGB-D saliency map quality evaluation,the segmentation quality,is designed,and the feature is used as the weight of the saliency map to perform linear fusion at the image level.At the pixel level,considering the differences of the significant values of a single pixel in different saliency maps,the saliency similarity of a single pixel is calculated,which is taken as the weight of the pixel and fused at the pixel level.Finally,the saliency maps obtained at these two levels are fused in proportion to obtain the final image saliency map.In this paper,a comparative experiment was carried out on the public dataset rgb-d1000.The experimental results show that the effectiveness of these two RGB-D saliency map optimization methods are better than the original image saliency detection method,and the saliency map fusion algorithm proposed in this paper has certain advantages compared with other fusion methods.