ADAPT Researchers, Dr. Matej Ulicny (Trinity College Dublin), Prof. Vladimir Krylov (Dublin City University) and Prof. Rozenn Dahyot (Maynooth University) have co-authored a paper titled “Harmonic convolutional networks based on discrete cosine transform”. This paper has been published in Pattern Recognition, a prestigious journal which underpins developments in cognate fields such as computer vision, image processing, text and document analysis and neural networks.
The paper proposes learning filters as combinations of preset spectral filters defined by the Discrete Cosine Transform (DCT), similarly to how Convolutional Neural Networks (CNNs) learn filters in order to capture local correlation patterns in feature space. In the paper, the researchers propose DCT-based harmonic blocks to replace conventional convolutional layers to produce partially or fully harmonic versions of new or existing CNN architectures. For more information, you can access the paper here.
