Shin Fujieda, Kohei Takayama, Toshiya Hachisuka. 2018.

https://doi.org/10.48550/arXiv.1805.08620

Wavelet CNN

• Transform the image with wavelet analysis (the paper used Haar wavelets.)
• High and low frequency outputs are routed between layers of convolution and bypass layers so gradients can pass through the layers.

CNNs learn patterns in the input but do not appear to be good at solving multi-resolution tasks; wavelets seem to give the CNNs access to spectral data and the frequency layers fed to CNNs so they can also reason about multiple layers.

Overfitting protection

• Images are scaled to a set size to be fed to the network
• An even smaller section is used for classifying; when training the network images have a random offset and rotation to generate variations to fight against overfitting.