---
tags: Image processing
---
# Robust Reflection Removal with Reflection-free Flash-only Cues
A simple yet effective reflection-free cue for robust reflection removal from a pair of flash and ambient (no-flash) images.

## Reflection Removal
區分真實的物件與反光
共有兩種光源
1. 自然環境光 ambient light
2. 相機閃光 flash


Ambient illumination (RGB/raw images under ambient illumination.)
$$
I^{raw}_a = R^{raw}_a+T^{raw}_a
$$

Flash illumination (RGB/raw images under flash illumination.)
$$
I^{raw}_f = R^{raw}_a+T^{raw}_a+R^{raw}_{fo}+T^{raw}_{fo}
$$

Flash-only illumination

$$
I^{raw}_{f0} = I^{raw}_f - I^{raw}_a = R^{raw}_{fo}+T^{raw}_{fo}
$$

> reflections of ambient image are invisible in the flash-only image $R^{raw}_{a}\approx 0$

## Contribution
- the reflection-free flash-only cue
> This cue is robust since it is independent of the appearance and strength of reflection.
- We propose a dedicated framework that can avoid introducing artifacts of flash-only images while utilizing reflection-free cues.
- We construct the first dataset that contains both raw data and RGB data for flash-based reflection removal.

## Architecture
估計反光
$$
\hat{R}_a=g_R(I_a,I^g_{fo};\theta_R)\\
L_R(R_a, \hat{R}_a)=\|R_a-\hat{R}_a\|^2_2
$$
估計去反光後的圖像
$$
\hat{T}_a=g_T(I_a,\hat{R}_a;\theta_T)\\
L_T(T_a, \hat{T}_a)=\|T_a-\hat{T}_a\|^2_2
$$

> ISP (Image Signal Processor)
> 把 Sensor 輸出的 Raw Data 進行處理，使之成為符合人眼感受的信號

## 補充
### Rendering equation
描述從 $x$ 點沿某一方向看的光放射的總額
$$
L_o(x,\vec{w})=L_e(x,\vec{w})+\int_{\Omega}f_r(x,\vec{w}',\vec{w})L_i(x,\vec{w}')(\vec{w}'\cdot\vec{n})d\vec{w}'
$$

| symbol | description |
|----------------|----------------------------------|
|$L_o(x,\vec{w})$|特定位置 $x$ 及角度 $\vec{w}$ 的出射光|
|$L_e(x,\vec{w})$| 同一位置及方向發出的光 |
|$\int_{\Omega}$ | 入射方向半球的無窮累加 |
|$f_r(x,\vec{w}',\vec{w})$| 該點從入射方向的出射方向的反射比例|
|$L_i(x,\vec{w}')$| 入射光位置及方向 $\vec{w}'$ |
|$(\vec{w}',\vec{n})$| 入射角帶來的入射光衰減 |

## 連結
[Paper](https://arxiv.org/pdf/2103.04273.pdf)
[Github](https://github.com/ChenyangLEI/flash-reflection-removal)
[ISP](https://iamard.blogspot.com/2014/01/isp-image-processing-pipeline-overview.html?view=flipcard)