Light field (LF) cameras can capture the scenes from different directions, which enables many applications, such as depth estimation and saliency detection. However, they can be seriously affected when there is poor luminance. This work focuses on LF image restoration under low-light conditions, which is a challenging task to deal with noise, color distortion and multiple views. We develop a learning-based decomposition-enhancement method to decouple this complex task into several sub-tasks. First, there is a lightweight DecomNet with spatial-angular residual modules to decompose each input LF image into its reflectance and illumination. Then, there is a reflectance EnhanceNet for noise suppression and color correction, and an illumination EnhanceNet for luminance recovery. They integrate a multi-scale pooling attention mechanism to extract more informative spatial and angular features, and a deep spatial-angular feature fusion group with selective fusion units to fully encode and fuse the spatial-angular information with high efficiency. Extensive experimental results have demonstrated that our method can restore the spatial details, geometric structures, luminance and color of LF images with various light levels effectively.

光场（LF）相机可以从不同方向捕捉场景，这使得许多应用成为可能，例如深度估计和显着性检测。但是，当亮度不佳时，它们会受到严重影响。这项工作侧重于低光照条件下的 LF 图像恢复，这是处理噪声、颜色失真和多视图的一项具有挑战性的任务。我们开发了一种基于学习的分解增强方法，将这个复杂的任务分解为几个子任务。首先，有一个带有空间角度残差模块的轻量级 DecomNet，可将每个输入 LF 图像分解为其反射率和照明度。然后，有一个用于噪声抑制和色彩校正的反射 EnhanceNet，以及一个用于亮度恢复的照明 EnhanceNet。他们集成了多尺度池化注意机制以提取更多信息空间和角度特征，以及具有选择性融合单元的深度空间-角度特征融合组，以高效地完全编码和融合空间-角度信息。大量的实验结果表明，我们的方法可以有效地恢复各种光照水平的低频图像的空间细节、几何结构、亮度和颜色。