dc.description.abstract | In three dimension digital city modeling, photographic fa?ade texture images are commonly attached to building faces in order to generate more realistic building models and scenes. However, there are usually occlusions from trees, cars and other foreign objects in close-ranged fa?ade photographs. Therefore, it is worthwhile to develop effective methods for efficiently correcting the occlusions in fa?ade texture images. Image inpainting is an effective approach to remove unwanted objects and fill holes in digital images. However, directly applying existing, general-purpose digital inpainting algorithms to the correction of fa?ade texture images with occlusions may not produce satisfactory results. This research developed a constrained image inpainting algorithm specifically designed for the correction of occluded fa?ade texture images. The objective is to remove selected occlusions of fa?ade texture images and restore the damaged texture blocks with reasonable textures identified with the developed constrained inpainting algorithm. The developed inpainting algorithm combines isophote, normal vector and confidence terms to calculate the best fill order of inpainting.
In comparison with existing inpainting approaches, it can restore the occluded texture blocks more reasonably based on fa?ade structures such as pillars, girders and window frames. A few parameters, such as window size and search area of inpainting were also investigated to better understand their effects on the inpainting results. Special constrains on the direction of inpainting order, such as top-down and bottom-up were also tested to understand their effects because the features of texture fa?ade were presumed to be along a particular direction. The developed constrained inpainting algorithms were applied to real building fa?ade images to valid their performance and to identify appropriate inpainting parameters for correcting fa?ade texture occlusions. Finally, the most widely used full-reference quality metric mean square error (MSE) and structural similarity index (SSIM) were employed to quantitatively evaluate the inpainting results.
Several complex building models are used to test the constrained image inpainting algorithms. Experimental results demonstrate that the proposed method based on fa?ade structures is effective in restoring occlusion of building fa?ade texture images. In addition, the experimental results of building fa?ade texture images with special structures are also validated, proving that the proposed approach can restore more reasonable visualization results.
| en_US |