Matrix
Norms
$$
\begin{gather}
\begin{aligned}
|| \mathcal{X} ||0 &= \sum{ i=1 }^{ n } \frac{ \mathcal{X}_i }{ \mathcal{X_i} }, \mathcal{X}i \neq 0 \\
|| \mathcal{X} ||p &= \Big[ \sum{ i = 1 }^{ n } \mathcal{X}i^p \Big]^{1/p} \\
|| \mathcal{X} ||{\infin} &= \mathcal{X}{max}
\end{aligned}
\end{gather}
$$
Mode-n Unfolding Matrix
沿着第$n$个纬度把张量展开
Global Pooling
| Category |
Concept |
| 全局平均池化(Global Average Pooling, GAP) |
对特征图上的所有元素取平均值 |
| 全局最大池化(Global Max Pooling, GMP) |
对特征图上的所有元素取最大值 |
Q-Shift
| Category |
Concept |
| 计算效率高 |
|
| 模型压缩 |
|
| 保留表达能力 |
|
指标介绍
MS-Fusion
| Category |
Concept |
| HS-HRI $\longrightarrow$ HS-LRI |
$\widehat{ MS }_k = (f * \widehat{ MS }_k) \downarrow^k + \pmb{N}_k$ |
| $f$ |
模糊算子 |
| $\pmb N$ |
噪声 |
| $\downarrow^k$ |
下采样$n$倍 |
| Category |
Concept |
| HS-HRI$\longrightarrow$ MS-HRI(Pan) |
$\pmb P = \pmb r^T \mathcal{MS} = \sum_{ k=1 }^{ N } \pmb{r_k MS_k}, \sum_{ k =1 }^{ N } r_i = 1$ |
Cross Correlation
CNN
Category
| Category |
Concept |
| Width |
太宽容易过拟合 |
| Depth |
过深，有些层可能会失效；梯度可能会消失 |
| Multi-Path |
|
| Feature-Maps |
卷积核对特征的提取 |
| Attention |
对图像重点部分的关注 |
| Channel(Dimension) |
分路卷积，分别提取Spatial/Channel Imformation |
| Channel Boosting |
特征增强 |
Application
Denoise
- AWNI(Additive White Noisy Imgae)
- Changing Network Architecture
| Category |
Concept |
| Fusion |
|
| 改变Width, Depth |
|
| 改变Loss Function |
|
| Cascade Connection |
|
| Skip Connection |
|
| Plugins |
Activation, Pooling, Fully-Connected Layer(展开成一维) |
-
Real Noisy Image (现实中的噪声)
-
Blind Denoising (盲源去噪)
Challenge
| Category |
Concept |
| Hyper-Para Settings |
超参数设置十分重要 |
| Hard to Explain |
|
速记
Important Networks
Mamba Network
Residual Network
Appendix
Concept
| Category |
Concept |
| Weak Edge-Information Noisy Image |
物体与背景的边缘模糊 |
| Sparse Input |
大部分输入数据为0 |