Image Segmentation In RGBD Via Saliency Seeds

As a key step in image analysis,image segmentation has become a hot topic in the field of image processing.Image segmentation is the process of dividing an image into non-overlapping and non-intersecting regions according to different features and extracting objects of interest.It has been widely used in many different fields such as military,medical,and transportation.However,due to the complexity and diversity of images,there is no mature image segmentation method that meets various requirements currently.Therefore,it shows great significance to the development of computer vision by deeply researching image segmentation technology and exploring accurate and efficient image segmentation methods.The interactive image segmentation algorithm incorporates the prior knowledge provided by the users,which can achieve satisfied segmentation results.This algorithm is the current mainstream image segmentation method.However,interactive segmentation relies heavily on the subjective experience of the users,which is time-consuming,labor-intensive,and does not have real-time nature.With the development of 3D sensing technology,image segmentation algorithms are no longer limited to two-dimensional images.RGBD images can provide richer three-dimensional scene information than color images.Therefore,scholars have focused on RGBD image segmentation technology.In view of the interactive image segmentation's shortcomings and the RGBD's advantages,an automatic image segmentation in RGBD via saliency seeds method is proposed.Superpixel segmentation preprocessing,saliency detection and region-based image segmentation algorithms are studied emphatically.In this paper,the improved image segmentation method is used to obtain good segmentation results.The main research contents are as follows:1.Propose a depth-weighted superpixel segmentation algorithm.Six-dimensional feature vectors based on color,position,and depth are constructed to obtain more accurate initial superpixel segmentation results.2.Propose a multi-feature fusion saliency seeds detection method.The images' color feature,depth information and center prior are fused in this algorithm.The foreground seeds and background seeds are detected by the hitting times.The hitting times is obtained by constructing global and local graph for RGBD images.This method can effectively improve the accuracy of saliency seeds detection.3.Propose a new automatic image segmentation method,which is a maximal similarity region merging algorithm that combines with saliency seeds.Firstly the image is pre-segmented by the depth-weighted superpixel segmentation method.Then the multi-feature fusion saliency seeds detection method is used to obtain the foreground seeds and background seeds.Finally the maximal similarity region merging mechanism is applied to segment the image,and using the foreground seeds and background seeds as interactive information.