IIIT Hyderabad Publications |
|||||||||
|
Building detection in VHR Satellite Imagery by using Fast ICA and MSERAuthor: LIPIKA AGARWAL Date: 2017-12-23 Report no: IIIT/TH/2017/91 Advisor:K S Rajan AbstractSatellite imagery in the recent past has gone through numerous advancement. Different satellite sensors have been developed with the capability to capture the data at different scales from coarse to very fine. The finer resolution data provides detailed information which helps in improving the object detection for more number of land cover classes. Satellite imagery forms the backbone of remote sensing. Remote sensing has a wide range of applications in Earth sciences like geology, military, economic and humanitarian purposes. Urban remote sensing is useful for analysing and planning. Building detection is one of the most important aspects of urban remote sensing. Building footprint detection is very useful in wide range of applications like disaster management, urban planning, cartography, environment modeling, all require knowledge of man made structures. Even now, mapping of buildings is done using remote sensing based on visual interpretation or manual digitizing of digital images. These methods are tedious, time consuming, requires qualified people, therefore it has virtually no scope for real-time applications. This work presents an approach to automatic building detection from very high resolution (VHR) satellite images. The work includes discussion of various image processing techniques used and challenges faced with detection of different building models. Monocular satellite data is used for this work. Satellite images typically contain elements like development areas, vegetation, water bodies, barren land etc. Development areas include buildings, roads, pavements and other such man-made entities. The first task of building detection process is to remove the natural entities and focus on development regions. This task is realized using Independent Component Analysis (ICA). Three input planes derived from the LAB and LUV colorspace are used as input to the ICA which in turn converts them into three independent output planes. One of these output planes primarily contains man-made structure pixels and other two primarily contain vegetation and other natural entities. Development areas are detected using pixels obtained from the plane and appropriate thresholding. The only drawback of this technique is buildings are obtained in the form of individual pixels and as a result, object-level performance of the detection is not as good as its pixel-level performance. The second phase of this work builds on the learning of ICA based building detection. The candidate building pixels obtained through ICA are now used to extract maximally stable extremal regions (MSER) which are then filtered using geometric properties to obtain final candidate buildings. The technique is aimed at improving object-level performance. The achieved Precision of 80.64% and Recall of 83.65 % in case of object-level evaluation. The work offers an unsupervised building detection technique which is robust towards size, shape, color, types of rooftops and shadows. A wide test image set consisting of 15 localities is used to evaluate the performance of the complete detection process. This technique achieved higher object-level performance compared to the only ICA-based technique. The ICA and MSER combined approach have potential to detect buildings under different scenarios, not compromising with accuracy. Full thesis: pdf Centre for Language Technologies Research Centre |
||||||||
Copyright © 2009 - IIIT Hyderabad. All Rights Reserved. |