2023-03-07, 15:00–15:30, Auditorium
An overview of deep internal learning methods applied for satellite imagery with examples of open-source PyTorch implementations.
This talk provides an outline of the efforts to explore the use of Deep Internal Learning Methods for processing of satellite images. Internal learning allows to solve various computer vision tasks without the need of pre-training on wider datasets and without using any pre-trained models. The task, such as in image inpainting, is completely solved by using the information contained in the input sample.
The talk will briefly summarize the past work done in the field of computer vision, followed by an overview of the works applying the related techniques to satellite images. This allows for performing operations such as image completion, image super-resolution, or image denoising without the need of acquiring any datasets or extensive training of the network.
Finally, the talk will include a demonstration of several open source implementations of Deep Internal Learning techniques by running the examples in jupyter notebooks. Software used for the demonstration will be based in PyTorch, PyTorch Lightning, Numpy.