High-resolution depth images form the basis for various imaging applications in science, technology, and industry. However, capturing them with specialized equipment is not always feasible. Instead, in monocular depth prediction, a depth image is reconstructed from a conventional reflectance image that is more readily obtainable. Recently, machine learning methods have emerged as a particularly effective way of achieving this task. However, the success of these methods heavily depends on the available training data. Again, it often takes a lot of effort to obtain suitable data. In this paper, we present a method to easily generate surrogate surface images. For that purpose, we introduce ENGinnSAND, a database comprising 5000 pairs of reflection images and corresponding depth images of line-like surface structures. Additionally, we develop a hybrid Radon transform network, where the characteristic features in the images are analyzed in Radon space, and evaluate it against other established network architectures. Both the database and the proposed network architecture can be useful in any application where the underlying dataset contains line-like structures.