Exploring the Application of Convolutional Neural Networks for Photogrammetric Image Processing
Close-range photogrammetry is widely used to measure the surface shape of various objects and its deformations. Usually, a stereo pair of images of the object under study, obtained from different angles by means of several digital video cameras, is used for this purpose. The surface shape is measured by triangulating a set of corresponding two-dimensional points from these images using a predetermined location of cameras relative to each other. Various algorithms are used to find these points. Several photogrammetric methods use cross-correlation for this purpose. This paper discusses the possibility of replacing the correlation algorithm with neural networks to determine offsets in the images. They allow to increase the calculation speed and the spatial resolution of the measurement results. To verify the possibility of their application, a series of experimental images of surfaces with different deformations were obtained. Computational experiments were performed to process these images using selected neural networks and a classical cross-correlation algorithm. The limitations on the use of the compared algorithms were determined and their error in restoring the three-dimensional shape of the surface was estimated. The physical simulation to verify the selected neural networks for image processing for the task of photogrammetry showed their performance and efficiency.