PairLIE論文閱讀筆記

論文為2023CVPR的Learning a Simple Low-light Image Enhancer from Paired Low-light Instances.論文鏈接如下：

openaccess.thecvf.com/content/CVPR2023/papers/Fu_Learning_a_Simple_Low-Light_Image_Enhancer_From_Paired_Low-Light_Instances_CVPR_2023_paper.pdf

出發點

1.However, collecting high-quality reference maps in real-world scenarios is time-consuming and expensive.

出發點1：在低光照領域，從現實世界中獲取高質量的參考照片進行監督學習，既費時又困難，成本昂貴。

因為獲得低光環境的照片是容易的，而此低光照片對應的亮度較大的參考圖片是難得的。

2.To tackle the issues of limited information in a single low-light image and the poor adaptability of handcrafted priors, we propose to leverage paired low-light instances to train the LIE network.

Additionally, twice-exposure images provide useful information for solving the LIE task. As a result, our solution can reduce the demand for handcrafted priors and improve the adaptability of the network.

出發點2：為了解決手動設置的先驗的低適應性，減少手動設置先驗的需求，同時提升模型對陌生環境的適應性。

創新點

The core insight of our approach is to sufficiently exploit priors from paired low-light images.

Those low-light image pairs share the same scene content but different illumination. Mathematically, Retinex decomposition with low-light image pairs can be expressed as:

創新點1：作者利用兩張低光圖片進行訓練，以充分提取低光圖片的信息。

instead of directly imposing the Retinex decomposition on original low-light images, we adopt a simple self-supervised mechanism to remove inappropriate features and implement the Retinex decomposition on the optimized image.

創新點2：作者基于Retinex理論，但是并不循舊地直接運用Retinex的分解。作者采用一個簡單的自監督機制以實現不合理特征的去除（通常是一些噪音）以及更好地實現Retinex理論。

模型

將兩張同一場景不同曝光的低光圖片送入訓練中，圖片I1與I2先經過P-Net去除噪音，得到i1與i2，然后利用L-Net與R-Net分解為照度L1與反射R1（對應有L2與R2）。

在測試，只需要輸入一張低光照圖片I，經過P-Net的噪音去除，得到i，然后用L-Net與R-Net分解為照度和反射，然后對照度L進行增強，操作為g(L)，把增強結果與反射R進行元素乘法，得到增強后的圖片Enhanced Image。

設計及其損失

Note that, this paper does not focus on designing modernistic network structures. L-Net and R-Net are very similar and simple,

1.模型使用的L-Net與R-Net十分簡單。整體架構只是單純的卷積神經網絡。

Apart from L-Net and R-Net, we introduce P-Net to remove inappropriate features from the original image. Specifically, the structure of the P-Net is identical to the R-Net.

2,P-Net被設計用于去除不合理特征。
$$L_p=\mid \mid I_1?i_1\mid {\mid }_2^2$$

Note that the projection loss needs to cooperate with the other constraints to avoid a trivial solution.i,e.,i1 = I1.

3.Projection Loss：最大程度限制去除不合理特征后的i1和原始低光圖片I1的區別。

這個損失需要避免一個特例，即降噪后圖片與原圖相同，即未降噪。
$$L_c=\mid \mid R_1?R_2\mid {\mid }_2^2\text{\hspace{1em}(1)}$$

Since sensor noise hidden in dark regions will be amplified when the contrast is improved.

In our method, the sensor noise can be implicitly removed by Eq. 1.

4.Reflection Loss：通常用傳感或攝影設備拍攝低光場景照片會攜帶一定的設備噪音，這個損失最大限度保證兩張圖片的反射是相同的，減少傳感或攝影設備的影響，這是因為圖片場景的內容相同。

這個損失是確保反射的一致性。
$$L_R=\mid \mid R°L?i\mid {\mid }_2^2+\mid \mid R?i/stopgrad(L)\mid {\mid }_2^2+\mid \mid L?L_0\mid {\mid }_2^2+\mid \mid ?L\mid {\mid }_1$$

$\mid \mid R°L?i\mid {\mid }_2^2$ is applied to ensure a reasonable decomposition.

$\mid \mid R?i/stopgrad(L)\mid {\mid }_2^2$ is to guide the decomposition.

Specifically, the initialized illumination L0 is calculated via the maximum of the R, G, and B channels：$L_0=\underset{c\in R,G,B}{\max }{I}^c(x).$

5.Retinex Loss：Retinex損失是為了限制分解組塊L-Net和R-Net以滿足Retinex的理論要求。

本文畢