Research Of Visual Saliency Computational Model On Remote Sensing Applications

Visual saliency is a key concept in the human visual attention mechanism,which provides an important basis for visual attention in spatial distribution.Based on this mechanism,the human visual system can effectively filter and extract the massive visual input information,and improve the visual perception efficiency.This paper focuses on the research of the application of visual saliency computational model in the field of remote sensing,aims to optimize the allocation of computational resources and improve the performance of traditional algorithms,and studies the saliency computational model based ROI(Region of Interest)extraction,multi-temporal remote sensing image registration and ship change detection technology in complex harbor background.The main work and innovation of this paper can be summarized as follows:1.Based on the adaptability analysis of existing saliency computational models in remote sensing application and the requirement of on-orbit real-time processing algorithm,this paper proposes a ROI extraction technology based on structural tensor saliency computational model.This method effectively overcomes the disadvantages of the traditional algorithms which are sensitive to factors such as the size,the spatial location and the foreground color difference of the ROIs.The ROIs in the full resolution saliency map calculated by this method realizes high contour retention,high recall rate and high target-to-background contrast.Aiming to further improve the computational efficiency of the algorithm and reduce the sensitivity to the high frequency interference information in the background,a ROI extraction method based on superpixel-to-pixel saliency analysis is proposed.The method improves the performance and computational efficiency of the proposed model through hierarchical saliency analysis.The experimental results of different spatial resolution remote sensing image datasets show that the method can meet the practical application requirements of ROI extraction for remote sensing image.2.Traditional image registration methods based on feature points matching have disadvantages in computational efficiency for multi-temporal large-scale remote sensing images.Aiming to solve this problem,a remote sensing image registration method based on compactness degree of saliency is proposed.The method takes the compactness degree of saliency as the prior measure,and unifies the feature extraction and the feature matching steps,which enhances the directivity of the feature extraction.This method can ensure lessperformance loss and effectively improve the computational efficiency of large-scale remote sensing image registration.In addition,in order to improve the efficiency of region-based multi-temporal remote sensing image registration,this paper proposes a hybrid model of remote sensing image registration based on salient region feature.Based on salient regional feature matching and the similarity measure function optimization,our model introduces a method of estimating the initial values of the affine transformation parameters by the matching regions to ensure the fast convergence of the similarity measure function.3.In the research of harbor area ship target change detection,based on the local context and algorithm design requirements,a hierarchical sea-land segmentation algorithm based on region homogeneity is proposed.This method overcomes the interference of complex background through a hierarchical procedure,and realizes reasonable allocation of computing resources.In order to meet the application requirement of dynamic change monitoring in complex harbor area,an automated harbor area ship change detection method is proposed based on the research findings of ROI extraction and multi-temporal remote sensing image registration.The overall framework of the algorithm is based on the spatial-temporal saliency model which is inspired by the physiological hypothesis of the human visual system.The algorithm has a clear functional hierarchy,realizes reasonable allocation of the limited computational resources on the algorithm level,and makes up the algorithm deficiencies in efficiency and automation.The experimental results show that the proposed algorithm can effectively overcome the information interference caused by illumination difference,target shadow and complex background changes,and can gain promising objective performance which meets the practical application requirements.