Research On Shadow Detection Technology Based On High-resolution Remote Sensing Image

Compared with the traditional remote sensing images,high-resolution remote sensing images contain more details,and provide richer terrain information.In the meantime,the shadow may produce big influences for analysis of these high-resolution images.On one hand,the shadow in images can provide important information for the image spatial recovery and 3D applications.On the other hand,the existence of the shadow will influence the quality of images seriously.In order to remove the impact of shadow in image analysis,the shadow detection technology becomes a promising research topic in the field of remote sensing.Traditional shadow detection methods only consider the optical characteristics of the image and ignore the correlation among the pixels,which make the detection results exist noise,edge blur and other problems.Based on the color and brightness characteristics of high resolution remote sensing images and the correlation among adjacent pixels,the optical characteristics and spatial features are combined with the Extended Random Walker(ERW)model.Specifically,the Support Vector Machine(SVM)is first utilized to estimate the probability maps.In the SVM classification step,the training samples are automatically obtained according to an automatic high-resolution image shadow detection process.Finally,initial classification map is refined with the ERW model,which can simultaneously characterize the shadow property and spatial information in satellite images to further improve shadow detection accuracy.The main work of this dissertation is as follows:1.In order to implement the SVM based object detection without human interaction,this dissertation presents an algorithm to obtain the pixel-wise detection probability maps.Firstly,the luminance information of the image is extracted by transferring the original image into the Hue-Illuminant-Saturation space,and a shadow feature image is acquired by the calculating the spectral ratio in the HIS space.Then,the Otsu method is used to segment the image into shadow and no-shadow areas.Next,the morphological erosion and dilation operations are performed on the initial detection map.Only those pixels in the central area of the initial segments will selected as training samples for the SVM.Finally,by performing the SVM classification in the defined feature space,the pixel-wise detection probability maps can be obtained.2.In order to increase the accuracy of the SVM based shadow detection method,this dissertation presents an ERW based method to jointly exploit the shadow color information and spatial information.It is known that the results obtained by the traditional pixel-wise shadow detection methods suffer from noise since these methods do not consider the correlations among adjacent pixels.In order to solve this problem,the ERW models spatial information in a Graph model.Through optimizing this model,the pixel-wise detection map obtained by SVM could be refined.Experimental results performed on real data sets demonstrate that the detection accuracy could be increased dramatically with the ERW based spatial optimization.3.In order to implement the proposed method in real applications,this dissertation designs a software system for shadow detection based on SVM and ERW.The software system was programmed based on Visual Studio 2010 programming platform,MFC application framework and computer vision open source library OpenCV.The software system can achieve the reading and display of high-resolution images,the estimation of the shadow probability maps,the SVM and ERW based shadow detection of the images,the calculation of the overall accuracy and other functions.