Building Detection From High Resolution Remote Sensing Imagery Based On Visual Saliency And Deformable Part Model

High-resolution remote sensing image is an important source of spatial information in national economic construction,national defense security,etc..The type of ground objects in the image is complex and detailed.Automatic detection of ground objects has always been the focus of high-resolution remote sensing image application.Among them,buildings,as a major part of a city,are also typical artificial objects with remarkable characteristics on high-resolution remote sensing images.Automatic identification and extraction of buildings cover many fields,such as mapping,urban planning,environmental monitoring,and data acquisition of geographic information system.It is of great significance for national defense,economic construction and social development.Based on the cognitive computing theory of visual information,this paper combines human visual perception and cognitive computing models with high-resolution remote sensing information extraction and image understanding.By simulating the process of human cognitive image object,this paper attempts to proceed from the perspective of cognitive psychology.Combined with visual attention,scene perception,perception organization and other aspects of visual cognition theory,the key technologies such as feature modeling,visual attention,scene perception,and object recognition based on visual perception organization on high-resolution remote sensing images are studied.In addition,a set of methods for extracting buildings from high-resolution remote sensing images is designed,based on simulating human visual cognition,and the object recognition accuracy and the degree of automatic translation are improved.The main research work in this paper includes:Firstly,this paper analyzes the current situation of the research on the object detection of buildings in high-resolution remote sensing image,and puts forward the visual cognitive computational model into the object detection of buildings.A set of research schemes to detect buildings from the whole scene of high-resolution remote sensing image is presented.The visual attention guides to the selection of significant areas the solution to solve the "where" problem,and the scene rapid perception provides a priori information to the judgment of overall cognition of object,so that the algorithm can solve the "what" problem.In the fifth chapter,a comprehensive experiment of building detection based on the proposed scheme is carried out on the whole scene of high-resolution remote sensing image,and remarkable results areachieved.Secondly,the physiological structure of human visual perception system and the basic principle of information processing are studied,as well as the selective attention mechanism based on visual perception.By analyzing the salient features of artificial objects based on buildings of high-resolution remote sensing image,the structure texture is presented as a salient feature of high-resolution remote sensing images,and a visual selection attention model based on structure texture is presented.The extraction of significant artificial areas on high-resolution remote sensing images was completed.Thirdly,based on the biological vision mechanism of human rapid scene perception,the significance of scene classification before object detection is analyzed.This paper summarizes the existing methods of scene classification,and studies the methods of scene feature extraction and feature analysis of high-resolution remote sensing images,as well as the method of scene classification of significant regions,using visual cognitive computing theory.This paper puts forward a method to classify the scene features of scene Gist based on the structure texture and the visual attention model to provide the basis for the object detection of high-resolution remote sensing images.Finally,this paper analyzes the feasibility of using visual organization rules to form building objects on high-resolution remote sensing images,and studies the multi-scale characteristics of building objects' overall structure and the modeling methods of building objects' spatial structures based on visual topology based on the spectral characteristics and geometric structures of buildings.A high-resolution remote sensing image building detection method based on deformable component model is proposed.