基于双注意力机制的雾天图像清晰化算法研究

Research on hazy image sharpening algorithm based on dual attention mechanism

针对传统去雾算法容易依赖先验知识以及恢复出来的清晰图像会产生颜色失真等问题,本文提出一种基于双注意力机制的雾天图像清晰化算法。首先将雾图输入编码器,经过下采样后得到特征图像;特征提取模块将多个特征提取基本块联结在一起,每个基本块由局部残差学习和特征注意模块组成,提高图像质量以及图像特征信息的利用率,增加网络训练的稳定性;然后通过通道注意力与多尺度空间注意力并行的结构处理特征图像,使得网络更加关注细节特征,提取更多关键信息,同时提高网络效率;最后将融合后的特征图像输入解码器中,经过多级映射,得到与输入大小匹配的雾密度图。实验结果表明,不论是对合成雾天图像或者真实雾天图像,本文算法能够高效地进行去雾处理,得到更自然的清晰图像。

Aiming at the problem that the traditional dehazing algorithm is easy to rely on prior knowledge and the restored clear image produces color distortion,this paper proposes a haze image clarity algorithm based on dual attention mechanism.Firstly,the haze image is input into the encoder,and the feature image is obtained after down-sampling.The feature extraction module connects several basic blocks of feature extraction together,and each basic block is composed of local residual learning and feature attention module,which improves the image quality and the utilization rate of image feature information,and increases the stability of network training.Then feature images are processed by parallel structure of channel attention and multi-scale spatial attention,which makes the network pay more attention to detail features,extract more key information,and improve network efficiency.Finally,the fused feature image is input into the decoder,and the haze density image matching the input size is obtained by multi-level mapping.The experimental results show that the algorithm presented in this paper can effectively remove haze and get more natural and clear images no matter for synthetic or real hazy images.