Orthoimage Super-Resolution via Deep Convolutional Neural Networks

Using high resolution (HR) images collected from UAV, aerial craft or satellites is a research hotspot in the field forest areas analyzing. In practice, HR images are available for a small number of regions, while for the rest, the maximum density various around 1 px/m. HR image reconstruction is a well-known problem in computer vision. Recently, deep learning algorithms have achieved great success in image processing, so we have introduced them into the field of processing orthoimages. At the same time, we noticed that orthoimages generally have colorful blocks of different sizes. Taking into account this feature, we did not apply the classical algorithms directly, but made some improvements. Experiments show that the effect of proposed method is equivalent to the effect of classical algorithms, however, at the preprocessing stage, it significantly saves time. An approach to the forest areas analyzing, including image segmentation and the tree spices classification is proposed. The results of numerical calculations are presented.