Research On Superpixel Segmentation Method Based On Perceptual Prior

With the development of technology,the size of the image is getting larger and larger.Therefore,the performance of image processing algorithm based on pixel is more and more unable to meet people's requirements.Super pixel segmentation algorithm is to deal with this problem.Super pixel is a pixel block with certain visual significance.After the image is divided into super pixels,the complexity of subsequent processing is greatly reduced,and the super pixels also have some local information and structure information.In the tasks of image segmentation,target tracking,human pose estimation and image scene analysis,superpixel segmentation is an effective preprocessing method.In order to solve the problem that the simple linear iterative superpixel can not effectively segment the weak edge of the object in the image,this paper studies the prior of contour closure perception in the original algorithm,and proposes a superpixel segmentation algorithm based on contour closure perception.This algorithm constructs a new distance formula by transforming edge spectrum into pixel level information such as pixel similarity and gradient map.The edge spectrum of image is obtained by edge extraction algorithm,then the region similarity matrix is obtained by Region scanning algorithm,and then the pixel similarity is calculated.At the same time,the distance transform spectrum of the edge spectrum and the gradient spectrum of the distance transform spectrum are obtained.Combined with pixel similarity and gradient graph,this paper improves the simple linear iterative clustering algorithm.Experiments on BSDS500 and PASCAL VOC 2012 data sets show that the algorithm can segment the weak edge in the image and improve the index of boundary recall.In view of the error between the real edge and the superpixel edge generated by the superpixel segmentation algorithm based on geodesic distance,an improved distance formula is studied.A superpixel segmentation algorithm based on forgetting perception and a superpixel segmentation algorithm based on feature aggregation are proposed.The algorithm based on forgetting perception adds forgetting coefficient,edge term and Euclidean distance from pixel to seed.It extracts the edge spectrum of the image,generates seed points according to the image,searches and allocates the pixel points according to the four neighborhood from the seed points,and allocates the pixel points at the edge to the seed points with the smallest distance according to the distance until all the pixels are allocated.The algorithm based on feature aggregation is to aggregate the features of all the pixels in the path,and calculate the distance between the seed point and the pixel point using the Euclidean distance.The seed points are generated,and the pixel points are searched and allocated from the seed points.When a pixel point is searched,the aggregated feature of the point is calculated according to the search path,and the distance is calculated with the new feature.It is also tested on BSDS500 and PASCAL VOC 2012 datasets.Experiments show that the above two algorithms can not only maintain certain superpixel regularity,but also accurately segment the edge,and achieve good results in various evaluation indexes.