Research On Classification For Remote Sensing Image Based On Superpixel

At present,the image classification technology has been very mature,but the final classification effect is not particularly promising for remote sensing images with a great variety of features and complex contents.To analyze the characteristics of visible remote sensing images object,this paper proposed an efficient method based on the superpixel classification algorithm to classify the remote sensing images,which has achieved excellent results and great practical significance.In this paper,the training image is segmented by SLIC algorithm to obtain super pixel segmentation result.Then extracting the color and texture visual features of superpixel samples put classifier to train and generate the classifier model.The visual features of the test images superpixel blocks are fed into the classifier model for classification,so that each superpixel block gets a category information.By combining the superpixel blocks of the same category,the whole remote sensing image is segmented and identified.The specific research of this paper includes the following three aspects:This paper introduced the improved texture information into original SLIC segmentation algorithm to form 11-D SLIC superpixel segmentation algorithm.Firstly,the de-mean gray images and gradient images are combined to enhance texture information and reduce noise interference,and we can extract the improved 6-D texture information from it.Then,the 6-D texture information is integrated into the original 5-D SLIC algorithm to form the improved 11-D SLIC super-pixel segmentation algorithm.This algorithm can make the shape of super-pixel more regular,and benefic for feature extraction and classification.In addition,the process object of 11-D SLIC super-pixel segmentation algorithm is Gaussian filtered remote sensing image which reduces the isolated small areas.In this paper,we improve the Gabor feature extraction method and increase dimensionality of Gabor features based on this improved 11-D SLIC segmentation results.This method can construct minimum bounding rectangle of superpixel,and fixedly divide rectangle area,then extract more accurate texture features of superpixel block.At the same time,this method also can adequately consider the spatial information of superpixels,and describe the typical characteristics of different objects more detailed.The experimental results show that the algorithm we proposed can improve the segmentation effect of remote sensing images.To validate the effectiveness of the proposed algorithm,we carry out the contrast experiment on classification algorithms of different features.The results show that the classification algorithm of fusion color and improved Gabor texture visual features has the better classification accuracy which can achieve 89.75%,and effectively improves the classification accuracy of remote sensing images.