基于Transformer的U型低照度图像增强算法

U-Shaped Low-Illumination Image Enhancement Algorithm Based on Transformer

摘要: 针对非均匀光照环境下照度自适应算法的固有缺陷以及基于CNN的图像增强模型固有的卷积运算导致的感受野受限、无法建立长距离的全局依赖等问题，本文提出一种融合CA-Transformer模块和全局注意力融合模块的U型网络RT-UNet。本研究设计了基于轴向多头自注意力机制的CA-Transformer模块作为特征提取和重建的基础模块。该模块在兼顾CNN与Transformer结构优点的同时极大地减少了计算复杂度。进而，为了建立不同尺度，不同分辨率特征图之间的信息交互与融合，搭建了全局注意力融合模块来替代之前的残差连接，帮助网络关注到更感兴趣的区域，方便其学习到更有用，更精细的特征。实验结果表明，本算法在主客观评价指标上相比于近几年一些主流图像增强算法均具有很强的竞争力。

Abstract: To solve the inherent shortcoming of adaptive algorithms on illuminance in non-uniform illumination environments, the limited receptive field and the inability to establish long-distance global dependence caused by the inherent convolution operation of the CNN-based image enhancement models, this paper proposes a U-shaped network RT-UNet that integrates both CA-Transformer modules and global attention fusion modules. Specifically, a CA-Transformer module based on the axial multi-head self-attention mechanism is designed as the basic module for feature extraction and reconstruction. This module takes into account the advantages of CNN and Transformer structures while greatly reducing the computational complexity. To establish the information interaction and fusion between feature maps of different scales and different resolutions, a global attention fusion module is constructed to replace the previous residual connection, Such a module makes the network pay attention to the region of more interest and facilitates it to learn more useful and refined features. Finally, experimental results show that the proposed algorithm has strong competitiveness compared with some mainstream image enhancement algorithms in recent years in terms of subjective and objective evaluation indexes.