Research On RGB-D Image Segmentation Algorithm Based On Contour Information

Image segmentation is one of the key steps to identify objects in the scene.In this paper,the following three aspects are studied based on the RGB-D images for the edge operator's vulnerability to noise and the over-segmentation and under-segmentation that are easily generated in the image segmentation process.1.An image edge detection algorithm is proposed based on RGB-D images' contour cues fusion to solve the problem of high detection error rate at the similar neighbourhoods in images.The algorithm uses multi-scales,multi-directions methods to respectively extract brightness contour cues,color contour cues,texture contour cues of RGB image,and depth contour cues and surface normal vector contour cues of depth image,and fuse all contour cues.Simulation experiments show that the edge detection algorithm with multiple cues fusion has good anti-noise performance.2.A weighted guided filtering algorithm for contour protection is proposed for the problem that the egde detection algorithm is affected by the background noise of the RGB image.Using the gradient of the pixel as the representation of contour information,a weighted model is introduced to propose a weighted guided filtering model.The model uses the contour information of RGB-D image to adjust the constraint factor and constrain the range of influence of the weight function so that the filter makes a more accurate response to the image contour.Simulation experiments show that the algorithm enhances the contour information of RGB image and filters out some background noise,improving the accuracy of edge detection.3.A hierarchical segmentation algorithm is proposed for RGB-D images based on region projection.The algorithm uses the oriented watershed transformation to transform the contour information of RGB-D images into region information,obtains the initial closed image segmentation,and converts it into an ultrametric contour map.Finally,the region projection algorithm is used to optimize the region segmentation result.Simulation experiments show that this algorithm can further reduce the probability of over-segmentation and under-segmentation,and improve the rate of the best segmentation covering.The NYUD2 dataset is used for experiments and the results are evaluated using the P-R framework and regional coverage guidelines.Experiment results show that RGB-D edge detection algorithm can effectively improve the average accuracy of the regional edge detection by contour cues fusion and using weighted guided filtering,and the projection-corrected hierarchical segmentation algorithm can effectively improve the rate of the best segmentation covering.