结合多尺度深度学习网络和Retinex理论的低光照图像增强算法

针对低光照增强任务缺乏参考图像及现有算法存在的色彩失真、纹理丢失、细节模糊、真值图像获取难等问题，本文提出了一种基于Retinex理论与注意力机制的多尺度加权特征低光照图像增强算法。该算法通过基于Unet架构的特征提取模块对低光照图像进行多尺度的特征提取，生成高维度的多尺度特征图；建立注意力机制模块凸显对增强图像有利的不同尺度的特征信息，得到加权的高维特征图；最后反射估计模块中利用Retinex理论建立网络模型，通过高维特征图生成最终的增强图像。设计了一个端到端的网络架构并利用一组自正则损失函数对网络模型进行约束，摆脱了参考图像的约束，实现了无监督学习。最终实验结果表明本文算法在增强图像的对比度与清晰度的同时维持了较高的图像细节与纹理，具有良好的视觉效果，能够有效增强低光照图像，视觉质量得到较大改善；并与其他多种增强算法相比，客观指标PSNR和SSIM得到了提高。

Low-light Image Enhancement Algorithm Combining Multi-Scale Deep Learning Networks and Retinex Theory

? ?This paper proposes a multi-scale weighted feature low light image enhancement algorithm based on Retinex theory and attention mechanism， aiming at the lack of reference images for low light enhancement tasks and the problems of color distortion， texture loss， detail blur， and difficulty in obtaining truth images in existing algorithms. In this algorithm， the feature extraction module based on UNET architecture performs multi-scale feature extraction on low illumination images to generate high-dimensional multi-scale feature maps； the attention mechanism module is established to highlight the feature information of different scales beneficial to the enhanced image and obtain a weighted high-dimensional feature map； finally， in the reflection estimation module， the network model is established by using Retinex theory， and the final enhanced image is generated by the high-dimensional feature map. An end-to-end network architecture is designed， and a group of self regular loss functions is used to constrain the network model， so as to get rid of the constraints of reference images and realize unsupervised learning. The final experimental results show that this algorithm can enhance the contrast and clarity of the image while maintaining high image detail and texture， and has good visual effect. Our algorithm can effectively enhance the low illumination image and greatly improve its visual quality.