| Abstract [eng] |
The goal of this work is to investigate and experiment with advanced deep learning algorithms for the automatic segmentation of satellite and aerial images, aiming to create accurate buildings contours. The research focuses on semantic segmentation methods, particularly applying convolutional neural networks (CNNs), so that each image pixel is accurately assigned to one of two classes: building or non-building. Experiments and research are conducted using the Python programming language, PyTorch framework and deep learning libraries based on it, which provide access to modern model architectures and facilitate their application. During the work, the aim is to evaluate the accuracy and speed of different algorithms in solving this specific satellite image segmentation task and to search for effective ways to optimize them. The need for such automated segmentation algorithms is significant – they allow for the efficient updating of maps, assist in planning urban development and infrastructure (e.g., road requirements), and can be used to monitor illegal construction. To achieve the best results and innovative solutions, this work will investigate, compare, and apply various contemporary model types, including widely used encoder-decoder architectures as well as newer, promising transformer-based models adapted for image segmentation tasks. |
