Harmonic convolutional networks based on discrete cosine transform

Matej Ulicny, Vladimir A. Krylov, Rozenn Dahyot

Abstract: Highlights•The harmonic block is designed to learn filter weights in the DCT domain.•Harmonic CNNs are constructed by replacing the convolutional layers.•Parameter learning in frequency domain improves performance.•High-frequency parameter truncation can efficiently compress new or trained CNNs.•The hamonic block can make a CNN invariant to illumination changes.