Retinex-ADNet:一个低光照图像增强系统

Retinex-ADNet:a Low-light Image Enhancement System

在光照条件不充足的情况下,拍摄的图像质量较差.对于低光照图像增强,基于Retinex的方法大多忽略了降噪.本文结合Retinex和卷积神经网络构建了一个有效的模型,包括3个子模块,分解模块、注意力降噪模块和亮度调整模块.分解模块用残差连接方式和空洞卷积来构建,以减少在分解过程中细节信息的丢失,得到更准确的反射图和亮度图.降噪模块引入了注意力机制,对亮度图处理后得到注意力图用来指导反射图的降噪.亮度调整模块对亮度图进行亮度调整.在不同曝光条件下拍摄的成对图像上训练模型,本文提出的Retinex-ADNet模型取得了更好的效果.

In the case of insufficient lighting,the quality of captured images is quite poor.For the enhancement of low-light images,Retinex-based methods mostly ignore denoising.This paper combines Retinex and convolutional neural network to build an effective model,including three sub-modules,decomposition module,attention denoising module and illumination adjustment module.The decomposition module is constructed with residual connection and dilation convolution to reduce the loss of detailed information during this process,and obtain more accurate reflectance maps and illumination maps.The denoising module introduces an attention mechanism,and the attention map is obtained after processing the illumination map to guide the denoising of the reflection map.The illumination adjustment module adjusts the brightness of the illumination map.By training the model on the paired image taken under different exposure conditions,the Retinex-ADNet proposed in this paper has achieved better results.