1. Motivation and framework
當前基於CNN的高光譜圖像修復取得了非常大的進展，但是仍然存在如下兩個問題：
- the trained model is limited to the model-driven noise simulation process and may generalize poorly to the real HSI with more sophisticated noise distributions.
- the whole-band approach may lead to an undesirable effect on the noise-free bands when only a few bands are degraded. In general, the noise degradation only occurs in a minority of the spectral bands, primarily due to the low SNR caused by the optical dispersion characteristics of the spectrograph. The noise intensity are varying among different bands, so that the whole-band assumption is not practical.
為了解決這兩個問題，作者提出了一個 self-supervised 降噪網絡（如上圖），包括 noise estimator 和 CNN denoiser 兩個部分：
- noise estimator的主要作用是 exploit the information of both the noise-free bands and the noisy bands，從而提取真實的噪聲樣本。 estimated noise samples and the clean bands from one single HSI are combined to build the paired training data. (noise sample 和 clean patches 都是從含噪聲的HSI中生成，沒有使用外部數據)
- CNN denoiser 是一個 multi-to-single 的 CNN網絡。noisy bands and adjacent bands are jointly aggreated via the contextual dilated spatial blocks and the spectral convolutional blocks.
論文主要有三個貢獻：
- 設計了自監督的網絡去除HSI的混合噪聲
- 設計了自監督的訓練方法（noise-clean image pairs通過多次回歸來從 noisy 高光譜圖像自身提取）
- 構建multi-to-single CNN去噪模型用於提取 spectral-sptial 特征
2. Noise estimator
作者將 multiple regression 和 LMLSD 集成到一起。（這個地方沒大看明白）通過這個操作，可以根據噪聲強度閾值把全部波段分為 clean bands 和 noisy bands。然后，成對的訓練數據是在 clean 數據上加上隨機的 noise samples 來實現。
3. Self-supervised CNN denoiser
(1) self-supervised training scheme. 當前方法大多規定了Gaussian noise，strip noise, and mixed noise 模型來進行訓練，當噪聲分布不符合這些模型時，影響性能。根據噪聲估計的結果，noisy HSI can be divided into clean bands and noisy bands. 成對的訓練數據是在 clean 數據上加上隨機的 noise samples 來實現。這樣，it is effective for the model to remove the same noise in the noisy bands including specific bands. 作者指出，這樣的設計就避免了當前方法嚴重依賴訓練數據的不足，同時可以學習數據集中的真實噪聲分布。
(2) Multi-to-single band denoising network. 作者提出了一個 multi-to-single 去噪網絡，如下圖所示。當前波段和相鄰波段為一組輸入\(\{y_{b-1}, y_b, y_{b+1}\}\) ，首先使用 multi-scale dilated conv 處理各個波段。作者指出，the spatial size of various ground objects can be quite different, so that it is difficult to capture enough contextual information from a single-scale receptive field. 因此，使用 multi-scale dilated conv aggregate the contextural information. (其實就是 dilated conv 應用了不同的步長) 。后面的 spectral denosing network 沒有太多特點，這里不多介紹。
(3) Spatial-spectral model training. 在 output 和 realistic noise sample 之間計算MSE損失，這里不過多介紹。
實驗部分可以參照作者論文，這里不過多介紹。