Research Of Image Segmentation Algorithm With Depth Information

Image segmentation is the technique that divide images into a number of areaswith specific and unique properties, however, color image is the2D projection of3Dscene, in the process of getting images, we lose depth information of the scene, soonly relying on color information, it is difficult to obtain ideal segmentation result.In this paper, we proposed a new segmentation algorithm with depth informationon the basis of a brief introduction of traditional color image segmentation algorithms,depth map and stereoscopic vision technology, this algorithm mainly contains thepretreatment of depth map, the rough segmentation by SLIC superpixel segmentationalgorithm, the evaluation of image segmentation result based on gradient, the precisesegmentation and region merging based on depth information.Focusing on the problem that depth maps generated by Kinect device have noiseand holes, we proposed a new depth hole filling algorithm based on improvedbilateral filtering, by improving color constraint and adding a median filter, then therepair result is more precise and smooth.In view of the problem that traditional image segmentation algorithms based oncolor information have some mistakes, we proposed a new image segmentationalgorithm to accurately segment images with depth information on the basis of SLICsuperpixel image segmentation result, this part mainly contains two aspects: one is theevaluation of segmentation result generated by SLIC based on the characteristics ofgood borderlines and pixel clustering of depth map, and then confirm the wrongsuperpixel regions; another one is the precise segmentation of wrong superpixelregions based on the characteristics of grayscale clustering and color independence ofdepth map, so as to revise segmentation result.Aiming at over-segmentation during the segmentation process, we proposed anew merging algorithm based on depth and color information, in this step, wecalculate the characteristic values of each superpixel block with color and depth information, then generate a similarity table by characteristic values, merge superpixelregions according to the similarity table, so as to reduce over-segmentation and get agood segmentation results.