目录:
- TNRD区别于一般NRD
- MSND与TNRD的联系
- MSND用于高斯去噪及其推导
- MSND用于泊松去噪及其推导
- 实验
贡献/创新点
(1)Multi-scale procedure is useful, 尤其是heavy noise。
(2)描述了TNRD和MSND的广泛使用(通用poisson denoising)
(3)Extend the TNRD with a multi-scale strategy, 可以增大感受野。
TNRD相比于一般的nonlinear diffusion process的主要变化:
−−−−−TRAINABLE−−−−−
即:The TNRD framework is modeled by highlyparameterizedlinearfilter① as well as highlyparameterizedinfluencefunctions② .
①②,they are learned from training data through a loss based approach.
⎧⎩⎨⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪n0=I0ut=Proxgtreactionforce⎛⎝⎜⎜⎜⎜ut+1−⎛⎝⎜⎜⎜⎜∑i=1Nk(Kti)⊤ϕti(Ktiut+1)diffusionforce+ψt(ut+1,f)reactionforce⎞⎠⎟⎟⎟⎟⎞⎠⎟⎟⎟⎟t=1,⋯,T(2.1)
其中,image u∈RN ; Ki∈RN×N , highly sparse matrix, a set of linear fitlers;
ϕ(Kiu)∈RN : a set of influence function / flux function;
Proxgt : proximal mapping operation.
*****TNRD结构****
The reaction term(force) is designated to handle a certain smooth data term, while the term g(u,f) is used to handle the problem with a non-smooth data term.
Proxgt(u~)=minu∥u−u~∥222+gt(u,f)
TNRD⟹MSND (增加了缩放尺度)
∑l=0L∑i=1NkA⊤l(Ktl,i)⊤ϕtl,i(Ktl,iAlut−1),(3.1)
公式 (3.1) 为公式 (2.1) diffusion force的更新部分;
*****MSND结构****
MSND的完整求解公式
Al : Down-sampling the original image u to different sizes.
The parameters to be trained in stage t Θt={
{
{
θtl,i,ktl,i}Ll=1}Nki=1,λt} ;其主要难点在于计算 ∂ut∂Θt 和 ∂ut+1∂ut
⎧⎩⎨⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪ut=Proxgtreactionforce⎛⎝⎜⎜⎜⎜ut−1−⎛⎝⎜⎜⎜⎜∑l=0L∑i=1NkA⊤l(Ktl,i)⊤ϕtl,i(Ktl,iAlut−1)diff