Superpixels Generation Methods On Images

Serving as a key step for application of computer vision and image processing,the image over-segmentation,i.e,the superpixel generation is one of the important research topic.Superpixels generation of images means that locally clustering the pixels accord-ing their features(e.g.,color,texture and coordinates),such that the pixels in the same superpixel has identical or similar features.Compared to the pixels,superpixel contain-s more local information,and can adhere most of the boundaries of the object in the images,so serving as a higher processing unit,superpixel has been applied to a variety of tasks of image processing and computer vision,and it can significantly improve the processing efficiency.Due to the complexity of natural images,the features of each pixel will influence each other when we clustering the pixels to superpixel.Thus,how to control the signif-icance of each feature(e.g.,color,texture and coordinates)to generate the superpixel is the key issue in the superpixels segmentation problem.Considering that the propagation based superpixels generation algorithms can take more image information into consider when over-segment the images,this article mainly discuss the superpixels generation algorithm based on propagation based algorithms.However,traditional propagation based superpixels generation algorithm mainly based on geodesic distance or level set,which considering the image contents on the propagation path too much,therefore,they cannot adhere to the boundaries much well.In addition,as an important feature of im-ages,texture information rarely be applied to the superpixel generation problem.As the advancement of the depth image capturing camera,the RGB-D images have attracted increasing attention.How to generate superpixels based on the depth images has im-portant research significance,too.According to the above issues,the research of this thesis are performed from the following aspects:(1)As the geometric flow based superpixels generation algorithm cannot adhere to the weak boundaries well,we analysis the diffusion process concretely,and come to the conclusion that they considering the image contents on the path too much.So we propose a flooding based superpixels generation algorithm,and combine the various features of the pixels through a distance function,to measure the dissimilarity between seeds and pixels,and thus enhance the over-segmentation accuracy.(2)Texture is an important feature for image segmentation,we combine the texture features of images into the superpixels generation algorithm.As the boundaries of the texture map is very ambiguous,especially in the regions near the hard edges.We extract the main structure of images firstly,and train the weights for each feature in the distance function according to the local complexity of the images,to effectively balance the significance of each feature for the superpixels over-segmentation,and enhance the over-segmentation further.(3)The superpixels generation algorithm based on RGBD images is also stud-ied in this paper.As depth image contains more information than 2D images,and we can transform the 2D depth image to 3D point cloud,to obtain more plentiful 3D ge-ometrical information.As previous superpixels generation algorithm mainly based on Euclidean distance,one of the major problem of Euclidean distance based superpixels generation algorithm is that,it cannot adhere to the edges which has similar contents along it.We transform the over-segmentation problem on RGBD images to the 3D tri-angle mesh,and define a geodesic driven metric on this mesh to measure the similarity between seeds and pixels,as this metric can effectively combine the features on the meshes,so it can enhance the over-segmentation accuracy greatly.