Study Of Superpixel Segmentation Method Based On Clustering

With the massive increase of images data,traditional segmentation algorithms based on pixels for image processing are increasingly unable to meet the needs of daily processing.In order to be able to effectively process these massive image data,the concept of superpixels is proposed.A superpixel refers to a local area connected by a group of pixels that are consistent in perception,and the superpixels do not overlap with each other.Superpixels combine pixels with similar visual representations to effectively reduce information redundancy,which greatly reduces costs of calculation and storage,helping to improve subsequent processing efficiency.In recent years,a large number of superpixels generation algorithms have been proposed.Among them,the simple linear iterative clustering algorithm(SLIC)and the simple linear non-iterative clustering algorithm(SNIC)are the most widely used superpixel segmentation algorithms.But they still have room for improvement.On the one hand,it only relies on color features and spatial features to distinguish pixels,which results in the algorithm's weak ability to distinguish boundary points and insensitive to background details;on the other hand,the K-means algorithm they are based on has certain limitations.The segmentation performance is difficult to improve.In response to these two aspects,This thesis mainly did the following two tasks:(1)A Multifeature Non-iterative Superpixels Segmentation(MNSS)algorithm is proposed.First,the algorithm enhances the feature representation of pixels.The algorithm introduces color gradient features and morphological contour features to enhance the sensitivity of the algorithm to color changes in details and the ability to fit the edges of the target.Secondly,the algorithm performs category labeling from the cluster center to its four neighbors,and completes the superpixels segmentation in a non-iterative way.Finally,the algorithm is evaluated on the BSDS500 public data set.The experimental results show that the algorithm achieves the expected effect,and compared with other comparison algorithms,it shows good segmentation performance.(2)A super pixel segmentation algorithm based on local feature rescaling(LRe Lab?MNSS)is proposed.First,in view of the poor performance of the K-means-based superpixels segmentation algorithm for clusters with uneven sizes,a feature rescaling operation is proposed.The feature rescaling operation performs rescaling based on the density distribution of features,thereby obtaining data clusters of relatively uniform size.Secondly,two feature rescaling methods are derived according to the different implementation methods: global feature rescaling method(Re Lab)and local feature rescaling method(LRe Lab),and experiments have verified that these two feature rescaling methods have an effect on K-means The effectiveness of algorithm improvements.Finally,based on the local feature rescaling operation,the LRe Lab?MNSS algorithm is proposed.The algorithm performance is compared on the BSDS500 public data set.The results show that the proposed LRe Lab?MNSS algorithm has better segmentation performance when the time complexity is similar.