Research On Remote Sensing Images Change Detection Method Based On Visual Saliency Model

A series of global ecological environment problems are caused by human consumption of resources and destruction of environment in the process of rapid economic development.The remote sensing technology has become one of the important means of ecological environment monitoring due to its characteristics of continuity,real-time,wide coverage,free from terrain constraints,many means and large amount of information.Among various technologies of remote sensing,one import is the remote sensing images change detection.Remote sensing images change detection technology is a process of using multi-temporal remote sensing images and other auxiliary data that covering the same surface area to analyze and determine the changes of land,water,vegetation and other surface features.It has been widely used in civil and military fields,and has great significance for promoting the sustainable development of the global economy and human social activities.However,one notable problem of SAR images change detection is that it’s difficult to balance the macroscopicity of change semantics and the microcosmic of pixel level detection.To deal with this problem,we introduce the visual saliency model to extract the macroscopical change semantics information.Then we use the optimization method to iteratively solve the stable pixel level change detection results to achieve completely extract the change objects.Based on above work,the detection accuracy has been improved effectively..In this thesis,the specific work summarized as follows:1.A change detection method of SAR image based on frequency domain saliency model and multi-scale subspace optimization is proposed.Firstly,two-temporal images are used to generate the multi-pattern difference images,and the multi-scale frequency domain saliency detection is carried out to obtain a series of multi-scale salient difference images,so as to remove the noise and rough positioning of the changing area.Secondly,the structural element with different shapes and sizes is selected to expand the salient difference images to construct the multi-scale structure subspace to fit the SAR images.Finally,a fusion algorithm based on Nytr?m sampling spectrum clustering is used to reconstruct the spectrum clustering similarity matrix by optimizing the weight distribution,and the final change detection results are obtained.The experimental results show that the comprehensive detection accuracy of this method is effectively improved.2.A change detection method of SAR image based on neural network semantic segmentation model and local dynamic energy function optimization is proposed.Firstly,two-temporal images are used to generate the multi-pattern difference images and expands the data to construct the training and testing data set of semantic segmentation network.Secondly,the unary potential function of conditional random field(CRF)model is obtained by semantic segmentation of the difference image,the energy window information entropy of multi-pattern difference images is used as the pairwise potential function of the dynamic fusion CRF model,and the two together construct the remote sensing change energy function.Finally,the mean field approximation algorithm is used to achieve CRF model reasoning and change detection results.The experimental results show that our method has accurate semantics,clear boundary and higher detection accuracy than other methods.