Enhancing 3D Building Model Textures with Super-Resolution of Aerial Photographs

Abstract. We have applied a super-resolution technique to enhance the texture image quality of LOD2 building models. Specifically, we adopted SwinIR for upscaling low-resolution images. In order to achieve better results, several approaches for creating training data, consisting of pairs of low-resolution and high-resolution image were investigated. The results showed that training with low-resolution images created by downsampling high-resolution images by a factor of four and then applying blurring and noise improved the sharpness of building edge lines in super-resolution images. Training data with augmentation techniques, such as the use of random noise and random rotation, are proved to be effective in enhancing super-resolution images. Using the super-resolved images, LOD2 building models were created, and a subjective evaluation of the building roof texture quality was conducted. The results indicated that for the input images used in super-resolution, 87% of buildings from high-quality aerial photographs and 78% from lower-quality photographs were rated as having sharp edges without distortion. Even with limited training data, the developed method was able to achieve high-quality super-resolution, regardless of the input image quality, leading to improved texture quality in LOD2 building models.