Research On Fast Superpixel Image Segmentation Algorithm And Its Application

In recent years,with the rapid development of computer science and technique,the level of image resolution is gradually increased,and the traditional pixel-level image segmentation methods are time-consuming for these high-resolution images.Superpixel is a kind of image pre-segmentation techniques that divide an image into a large number of small areas of homogeneity.Compared with traditional pixel-level image segmentation approaches,the superpixel segmentation can not only maintain local features of image,but also improve the computational efficiency of subsequent algorithms.At present,superpixel segmentation has been successfully applied in object detection and recognition of images,image segmentation,image classification,etc.The technique has attracted wide attention of many researchers.Therefore,it is of great significance to study the superpixel segmentation algorithms for the development of image segmentation theory.Furthermore,we apply superpixel algorithms to the classification of remote sensing images,which plays an important role in the development of ground observation technology in China.Although many superpixel segmentation algorithms were proposed,most of them suffer from problems such as low contour accuracy,high computational complexity and limited practical applications.To solve these problems,this paper studies the fast superpixel image segmentation algorithm and applies it to remote sensing image classification.The main research work of this paper is summarized as follows:(1)After each iteration of simple linear iterative clustering(SLIC)algorithm for image superpixel segmentation,there will be abnormal points with a low similarity to the cluster center,which affect the stability of the clustering center and is not conducive to the improvement of the contour fit;in addition,there are redundant calculations in the clustering process of SLIC algorithm,which will reduce the operation speed of the algorithm.To solve these problems,a fast superpixel image segmentation algorithm is proposed.Firstly,the pixels with low similarity between the cluster and the cluster center are removed,and use the remaining pixels in the cluster to update the cluster center,thereby enhancing the stability of the cluster center.It is beneficial to improve the hit rate of superpixel contours.Secondly,the edge pixels of each superpixel is regarded as unstable pixels,and the non-edge pixels of the superpixels are regarded as stable pixels and the category of the stable pixels is kept unchanged.Iterative labelling of unstable pixels can effectively reduce the computational complexity of the algorithm.Experiments show that,compared with mainstream algorithms,this algorithm can effectively improve the fit of superpixel contours while reducing the running time of the algorithm.(2)The application of superpixel algorithm in the classification of high-resolution remote sensing images can improve the classification speed of remote sensing images.However,most of the classification algorithms based on superpixel still have the problems of high computational complexity and low classification accuracy.To solve these problems,a multi-scale convolutional neural network(MCNN)model based on fast superpixel algorithm is proposed.The model first uses a fast superpixel image segmentation algorithm to pre-segment the image,and selects the center point of the superpixel block as the extracted pixel point of the image block.Compared with the traditional sliding window point selection method,the redundant marked pixels are greatly reduced at the same time,it improves the speed of the subsequent classification algorithm.Secondly,it is proposed to extract multiple image blocks of different scales at the center point of the superpixel block to form a data set,to reduce the influence of the scale effect caused by a single scale on the classification results,which is beneficial to the improvement of classification accuracy.Finally,a multi-scale CNN model is proposed,which is constructed by a convolutional neural network with multiple paths,to fully extract the characteristics of multi-scale image blocks to solve the classification problem of complex ground environment,and realize the accuracy of remote sensing ground features classification.Experiments show that,compared with mainstream classification methods,this method can effectively improve classification accuracy and reduce classification time.