Petar Vračar (2010) Learning and recognition of façade's structural elements from digital images. MSc thesis.
We describe a method for learning and recognizing windows a basic structural elements of façades. We assume rectangular windows and our method is adapted to the segmentation of façades which are horizontaly and verticaly aligned. These limitations are not very restrictive since vast majority of façades have these properties for practical and eastetic reasons. The method begins with perspective rectification of the input image, which ensures that the edges of objects, parallel in space, remain parallel in the image plane. The rectified image is then segmented into hierarchical structure of window candidates. Four heuristic functions for estimating the quality of generated window candidates are presented. They are defined on the basis of position in the hierarchical structure, the distribution of horizontal and vertical edges, the similarities between candidates, and the learned window appearances. The generation of the final hypothesis uses a greedy approach, which favors homogeneous and aligned candidates. Experimental results are satisfactory but the method should be supplemented before the practical use. The main disadvantage of the method is ignorance of general background knowledge of façade structures. The proposed method can be used in the calculation of approximate solutions and as a starting point for methods, like stochastic grammar, which includes background knowledge on the structure of symmetrical and repeated patterns of façades.
Actions (login required)