Abstract:

When Retinex is applied to many scenarios, its constraints and parameters are limited by the model capacity. A low illumination image enhancement algorithm based on deep learning is proposed, and a new network architecture Retinex-UNet（RUNet） is constructed. The architecture includes image decomposition network and image enhancement network. Firstly, the Retinex-Net idea is adopted. The Convolutional Neural Network（CNN） is used to learn and decompose the image, and then the result is used as an input to the enhanced network to perform end-to-end training on the input image. The enhanced network builds a U-Net-based network architecture that enhances images of any size. Validation on public data sets（LOL, SID） shows that the RUNet method has improved in performance, especially the overall visual effect.

Key words: Retinex-Net, low illumination image, convolutional neural network, U-Net, RUNet

摘要：

针对Retinex应用于多种场景时，其约束和参数会受到模型容量限制的问题，提出了一种基于深度学习的低照度图像增强算法，并构建了新的网络架构Retinex-UNet（RUNet）。该架构包含图像分解网络与图像增强网络两部分，利用Retinex-Net网络思想，通过卷积神经网络（Convolutional Neural Network，CNN）学习并分解图像，将其结果作为增强网络的输入，对输入图像进行端对端训练。在增强网络中构建了基于U-Net的网络架构，其可对任意大小的图像进行增强。通过在公开数据集（LOL，SID）上验证表明，RUNet方法在效果上有所改进，尤其是整体视觉效果。

关键词: Retinex-Net, 低照度图像, 卷积神经网络, U-Net, RUNet