A Study On The Recognition Algorithm Of Suspected Illegal Buildings In Fixed-point Monitoring Mode

Illegal buildings are the main obstacle to the development of our country.The thesis focuses on the study of change detection and object classification which are two important steps of the recognition for suspected illegal buildings on fixed-point monitoring mode.The main research contents of this thesis are as follows:(1)A multi-feature fusion for describing change differences algorithm is proposed,which contains three parts: based on the SSIM and adaptive color transferring algorithm for eliminating the local weather differences between old and new phase image,to create a pixel difference image;based on the local gradient difference algorithm for extracting the structure differences of buildings between old and new phase image,to create a texture difference image;based on the z-score normalized weighed algorithm for integrating pixel and texture difference image,to create a final difference image.Experiments show that the proposed algorithm can effectively describe the change differences of buildings between old and new phase image,and filter most of the "pseudo changes" which are caused by weather,plants etc.(2)A fuzzy-space MRF for extracting change regions algorithm is proposed.A improved iterative threshold algorithm is used to initially segment the difference image;a change region extraction algorithm based on MRF is used to update the initial segmentation,to solve the problem that some of the change regions of buildings are broken;a change region extraction algorithm based on fuzzy-space MRF is used to improve the space energy function of traditional MRF,to refine the edge of change regions and prevent change regions of different objects from being merged.Experiments show that the proposed algorithm can effectively extract the change region of buildings between old and new phase image,and the accuracy rate is 95.81%.(3)A deep learning network for recognizing buildings is researched.Using the advantage of CNN’s high dimensional expression,a building recognition method based on Res Net-152 is used to handle the complex scene,such as buildings are shaded,the views of building are diverse and the structure of building is complex.Experiments show the proposed method can effectively recognize the suspected illegal building,with the accuracy rate of 92.08% in building dataset and 91.67% in background dataset.Due to the particularity of illegal building recognition,a suspected illegal building recognition with integrating deep learning is used to handle this problem.Experiments show that the proposed algorithm can effectively detect suspected illegal buildings,with the accuracy rate of 94.65% and the error rate of 3.59%.