Research On RGB Image Segmentation Method Incorporating Depth Information

Image segmentation is the basis for visually guided robots or manipulators to identify and locate targets.The accuracy of image segmentation determines the accuracy of target recognition and capture.Traditional image segmentation methods are mostly based on color features such as grayscale and texture.When segmenting adjacent or overlapping target images with similar color features(such as stacked workpiece images or complex scene images,etc.),image segmentation methods based on color features are difficult to obtain accurate segmentation results because the nature of RGB image is the mapping of a three-dimensional space scene to a two-dimensional image,depth information is lost in the process.To this end,this paper researches and implements an image segmentation method that fuses depth information.It mainly includes three parts: preprocessing of depth images,superpixel segmentation by fusing depth information,and semantically consistent superpixel merge based on multi-feature fusion graph theory.The main research results of this article are as follows:First,a deep image hole classification and repair method combining local edge features of color images is proposed.The small holes and noise in the depth image are filtered by a bilateral filtering algorithm,and the holes are classified into edgeless and edged types according to the presence or absence of edge characteristics in the color image area corresponding to the remaining large holes.Holes without local edge features are repaired by means of mean filling.For holes with local edge features,the holes are segmented according to the edge features,and then the divided sub-holes are repaired layer by layer from the outside to the inside.Experimental results show that the depth image hole repair results of this method are better than the comparison method in terms of root mean square error,structural similarity and peak signal-to-noise ratio.Secondly,a superpixel segmentation method(SLIC-D)is proposed.Pixel similarity distance metric is established based on the 8-D feature metric matrix that fuses depth information.The similarity distance metric and the iterative process of the SLIC algorithm are used to cluster pixels to complete superpixel segmentation.The experimental results show that the proposed superpixel segmentation method has ahigher segment boundary recall rate and a lower mis-segmentation ratio,especially when there are adjacent objects with similar colors in the image.Finally,a semantically consistent superpixel merging method based on multi-feature fusion graph theory is proposed.Energy functions including data items,smooth items,and memo items are established.Combining Euclidean distance,covariance matrix distance,and depth image boundary distance,a multi-feature fusion method of neighboring superpixel similarity measurement is proposed,and the smoothing term of the energy function is established based on this.The label term is introduced into the energy function to remove redundant labels.In the process of minimizing the energy function,semantically consistent superpixel merge is realized.The experimental results show that compared with several existing methods,the image segmentation method based on the fusion of depth information in this paper has higher segmentation accuracy,and is superior to the comparison method in terms of operating efficiency and memory efficiency.