Year: 2013
Type: dissertation
Abstract: Information extracted from aerial photographs has found applications in a wide\nrange of areas including urban planning, crop and forest management, disaster\nrelief, and climate modeling. At present, much of the extraction is still\nperformed by human experts, making the process slow, costly, and error prone.\nThe goal of this thesis is to develop methods for automatically extracting the\nlocations of objects such as roads, buildings, and trees directly from aerial\nimages.\n\nWe investigate the use of machine learning methods trained on aligned aerial\nimages and possibly outdated maps for labeling the pixels of an aerial image\nwith semantic labels. We show how deep neural networks implemented on modern\nGPUs can be used to efficiently learn highly discriminative image features. We\nthen introduce new loss functions for training neural networks that are\npartially robust to incomplete and poorly registered target maps. Finally, we\npropose two ways of improving the predictions of our system by introducing\nstructure into the outputs of the neural networks.\n\nWe evaluate our system on the largest and most-challenging road and building\ndetection datasets considered in the literature and show that it works reliably\nunder a wide variety of conditions. Furthermore, we are releasing the first\nlarge-scale road and building detection datasets to the public in order to\nfacilitate future comparisons with other methods.
Source: TSpace (University of Toronto)
Authors: Geoffrey E. Hinton, Volodymyr Mnih
Institution: University of Toronto
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2623331213