Research On Superpixel Segmentation Based On Weight Function

With the popularity of mobile phones and the continuous development of the Internet,the use of images has greatly increased,the number of pixels in images and the quality of images have gradually improved.How to process images more quickly has become an important research direction in the field of image processing.Due to the continuity of image information,there is a large amount of redundant information in the image,and many adjacent pixels have similar colors and textures.Image superpixel segmentation is an important method to improve image processing speed.It refers to adjacent similar pixels as a whole,named superpixels,and superpixels replace pixels as the basic unit of image processing,thereby improving the efficiency of image processing.Superpixel segmentation is often used as a pre-processing step for image processing and has a wide range of applications in life,industry,and technology.The early superpixel segmentation algorithms were solved by analogy to the minimum cut problem of the graph.These algorithms treat pixels as vertices of graphs.The weights of edges between vertices represent similarities between pixels.The greater the similarity,the greater the weight of the edges.They divide the pixels into multiple classes,so that the similarity of similar pixels is as large as possible,and the similarity of different types of pixels is as small as possible.This type of algorithm produces superpixels that fit the boundaries,but the shape ofthe superpixels is irregular.With the rise of clustering algorithms,many superpixel algorithms are based on clustering algorithms.The cluster-based superpixel segmentation algorithms treat superpixels as clusters of pixels.These algorithms first initialize some seed points as cduster centers,and then assign each pixel to the seed point that is most similar to it.Pixels belonging to the same seed point form a superpixel.In order to improve the efficiency of these algorithms,they limit the search range of each seed point.After each clustering iteration,the color and position of the seed points are updated to the color mean and center of gravity of superpixels,and then the updated seed points are reclustered,and the above process is repeated until superpixels converge.Finally,to eliminate small blocks of superpixels,the algorithm merges them with the most similar superpixels around them.Since the superpixel clustering algorithms use the same distance metric function for all pixels,although it can generate regular superpixels,the superpixel boundary fit is not very well,and these algorithms do not guarantee the superpixel internal connectivity.Aiming at the above problems,this paper designs a superpixel clustering algorithm based on weight function.First of all,this paper defines a weight function,which can determine the weight of the color distance term and the spatial distance term in the distance metric according to the probability that the pixel falls on the boundary and the distance between the pixel and the seed point,so that superpixels fit the borders in complex areas of the image and are regular in the flat areas of the image.In order to further improve superpixel segmentation results,this paper designs a dynamic search strategy,which expands the search range of seed points with expansion trend to improve the accuracy of superpixel segmentation.Finally,this paper designs a merging strategy,which separates the internal disconnected superpixel into multiple internal connected sub-superpixels,and then merges the small superpixel with the most similar superp:ixels around it to ensure the internal connectivity of the superpixel.The analysis of the experimental results shows that the superpixel algorithm proposed in this paper can produce superpixels with regular and high degree of boundary fit at a low time cost.Compared with other superpixel segmentation algorithms,the proposed algorithm has improved significantly on several commonly used superpixel evaluation metrics,and its subjective visual effects are also better.