Research On Image Segmentation Algorithm Based On Improved Superpixel And Spectral Clustering

Image segmentation is one of the most important and basic steps in image processing.It has an important influence on object recognition in natural scenes,edge detection,medical image analysis and object tracking,etc.The classical approaches of image segmentation are pixel-based which lacking of the spatial and organizational contact representation of pixels.With the development of all kinds of camera equipment,the resolution of images that need to be processed is getting higher and higher in many application.Thus,the processing efficiency of algorithm is low when using the traditional spectral clustering to segment the images.To reduce the number of the pixels in an image,we use the superpixel to represent the pixels of the image.Then,improving the efficiency of representation and segmentation of the whole algorithm.To solve the limitation of the state of the art superpixel segmentation based on DBSCAN clustering algorithm and the traditional NJW spectral clustering method,we propose a novel method for image segmentation which is based on superpixel and spectral clustering.The main work and achievements of this dissertation are as follows:(1)There is a problem that some superpixels cannot adhere to the boundaries of objects well when using the superpixel segmentation which based on DBSCAN clustering algorithm.Thus,we brings out a modified way to improved the performance in adherence of boundaries.By converting the image from the RGB color space to the CIELAB color space,and we are fully considering the spatial information of the superpixel to modified the similarity measurement function between the superpixels in the clustering stage and merging stage.At the same time,Our searching strategy in clustering stage is optimized to expand the searching space and get more candidate points are as many as possible without increasing the running time,thereby improving the performance in adherence of boundaries.(2)There are certain issues when using traditional spectral clustering for image segmentation.Such as a high complexity and use the Gaussian kernel function as the similarity measure will be sensitive to scaling factor.To overcome aforementioned problems,a novel spectral clustering method with modified superpixel for image segmentation is proposed.Firstly,using the improved superpixel segmentation algorithm proposed in this paper to segment the image into several superpixels,and then each superpixel is taken as a pixel to reduce the number of the data point in an image.Based on this,we use a kernel fuzzy similarity measure to construct the affinity matrix of the spectral clustering.The result of the simulation experiment shows that the method proposed in this paper visible improvement both in diminishing segmentation error,and also it has a higher efficiency.