Neural-based image preprocessing for photogrammetic 3D reconstruction

Photogrammetry is a technique for deriving three-dimensional (3D) information from two-dimensional (2D) images. Traditional photogrammetric workflows heavily rely on manual preprocessing steps such as image enhancement, feature extraction, and noise reduction. Various neural network architectures, including convolutional neural networks (CNNs) and generative adversarial networks (GANs), have been adapted to address specific challenges encountered in photogrammetric image preprocessing. This thesis presents a method to improve the preprocessing stage in the photogrammetry by using neural-based approach for enhancing the photogrammetric 3D reconstruction. The objective of this thesis is to remove highlights from models in order to improve their 3D reconstruction. For this purpose, we can divide the work into two macro sections: the generation of the photorealistic dataset and the image translation. The dataset was generated from a set of models, and each one has been rendered in such a way to obtain pairs of photos, with and without highlights, which represents respectively the input and the target for the image to image translation model. By using this approach it is possible to enhance and improve the 3D reconstructed model by mitigating the influence of specular highlights.