[1].参考开源库1:https://github.com/fbcotter/pytorch_wavelets
Cotter, Fergal. Uses of complex wavelets in deep convolutional neural networks. Diss. 2020.
[2].参考开源库2:https://github.com/brunobelloni/wran-sr-pytorch
[3]
F:\A\1.Code\3.code\yimo\wran-sr-pytorch
DWT (Discrete Wavelet Transform)有三个缺点:对零交叉敏感性(zero-crossings sensitive, 在零值点正负切换相抵消,小波系数小[1])、方向选择性差(poor directional selectivity,DWT是实数变换)[1,3]和不具备平移不变性(not shift-invariant,输入信号中的微小平移可能会极大地扰动小波系数)
复数小波(Complex Wavelets)可以克服DWT的不足。启发于Fourier transform,Fourier transform具备好的方向选择性和平移不变性得益于复数余弦基:
进一步调整小波基,