Shape Detection Of Remote Sensing Object Based On Deep Learning

Shape detection of remote sensing object plays an important role in various higher level geographic and environmental applications,such as disaster assessment and rescuing,3D-city modeling,urban changing detection,earth observation and cartography and so on.Most advanced remote sensing object segmentation methods with Convolutional Neural Network(CNN)generally perform well.However,either big inadvertent fusion of adjacent instances or small scattered islands could happen due to the characteristics of aerial images themselves.Moreover,the target borders on remote sensing images,such as buildings,are difficult to delineate with precise structures,and usually complex post-processing would be adopted to generate smooth shapes.In view of the above circumstances,this thesis puts forward two technical routes based on contour:(1)A global-local two-stage offset prediction network for the delineation of building boundaries is proposed.The model mainly consists of two modules,namely,global displacement and local refinement.The global module could quickly guide the contour to the vicinity of the target boundary that is conducive to the stability of the model.The local module could attract the polygons towards the outline more precisely.Given an input image,we can predict the global displacement that shifts the polygons roughly by designing a CNN operated on the whole region.After extracting local features of contour vertices,circular convolution layers are adopted to further refine the positions.Comprehensive evaluations on the three available datasets show the effectiveness of the proposed method and obtain excellent performances.(2)In a new perspective of physical vibration theory for the first time,a novel contour vibration network(CVNet)is advanced to deal with building boundary delineation.Different from the previous contour-based methods,the work principle of CVNet conforms to Newton’s second law of motion according to the force analysis of an infinitesimal string.Based on the force and motion principle of string vibration,the motion of contour is connected to internal/external forces of string,which are driven by the characteristics of image/object itself.By performing the infinitesimal analysis on the contour string,we build a spatial-temporal contour vibration model,which is mathematically reduced to second-order differential equation.Further,the contour vibration equation is dynamic with parameterized coefficients to be learnt from the current contour state.The contour changes are finally evolved in a progressive mode through a recursive computation on contour vibration equation.Both the polygon contour evolution and the model learning are encapsulated into an end-to-end close-looping network framework.To verify the proposed method,we conduct extensive experiments on three public datasets.The experiments validate the feasibility that takes physical vibration theory for contour extraction,and the experiment results also demonstrate that our CVNet is effective and can achieve the state-of-the-art performances for the polygon-based building contour extraction.