The urbanization is on the way with the progress of human society. The surface of the earth is changing rapidly due to the construction and demolish of buildings and infrastructures. It is an urgent task to get the geographical information changes efficiently, and the change detection task has a close relationship to urban planning and monitoring, early warning and rescue of nature disasters. During the past years, there are many researches carried out about change detection based on remote sensing images. Especially after the introduction of very high resolution imagery, the change detection task goes to fine scales. The pixel value based approaches is highly limited due to it's need for precise registration between two images acquired by different times. Moreover, recent methods are usually restricted to specific problems and not suitable for the practical application. And also, most of the research focused on binary detection problems. In this thesis, we investigate the change detection problem at the object level, and develop methods with the theory from computer vision and object recognition. We both study the binary detection problem and the multi-category classification problem. In this work, 1) We collect and build an expert labeled database with different change images. This data set makes it possible to carry out the parameter training and model evaluation. 2) We propose an approach based on interest points and local features, which involves interest points and feature extraction, training of statistical classifier. This approach is less sensitive to noises. To handle the different view angle problem, we proposed a new matching algorithm, which improves the result. 3) We proposed an approach based on term-document representation to solve the multi-category change detection and classification problem. This approach involves feature extraction and quantification, image pair representation and kernel selection for SVM classifier. Moreover, we proposed two approaches to incorporate spatial information between visual words based on second order visual word and pLSA model to extract visual topics.
修改评论