基于多残差和多重特征融合的去雾算法

Fog removal algorithm based on multiple residuals and multiple feature fusion

针对目前大多数图像去雾算法由于细节丢失导致去雾后的图像颜色失真,雾霾残留以及纹理细节模糊等问题,提出一种基于多残差和多重特征融合端到端的去雾算法。首先通过设计浅层特征提取模块,为深层网络提高丰富信息的特征图;其次设计多残差级联模块,提取多层次特征,帮助模型学习更加复杂的特征表示;然后设计局部-全局特征融合模块,捕获从最细微到最广泛的特征;最后设计结合残差注意力的跨层特征融合模块,避免上下采样后的细节缺失,更好地提取图像中的局部与全局信息特征。实验结果表明,所提算法在SOTS室内、室外测试集上峰值信噪比(PSNR)分别取得了33.12、31.07dB,结构相似性(SSIM)分别取得0.986、0.983,与当前大多数主流算法相比得到了明显的提升,且在合成雾图像和真实雾霾图像均取得了不错的去雾效果,复原图像细节更加清晰,更符合人类视觉感知。

To address the common issues in most existing image dehazing algorithms,such as color distortion,haze residue,and blurring of texture details due to the loss of fine details,a new end-to-end dehazing algorithm based on multi-residual and multi-feature fusion is proposed.Initially,a shallow feature extraction module is designed to provide the deep network with feature maps rich in information.Subsequently,a multi-residual cascading module is constructed to extract multi-level features,assisting the model in learning more complex feature representations.Furthermore,a local-global feature fusion module is introduced to capture features ranging from the most subtle to the most extensive.Finally,a cross-layer feature fusion module,combined with residual attention,is designed to prevent the loss of details after upsampling and downsampling,thus better extracting local and global information features from the image.Experimental results show that the proposed algorithm achieves peak signal-to-noise ratio(PSNR)of 33.12 and 31.07 dB,and structural similarity(SSIM)of 0.986 and 0.983,respectively,on indoor and outdoor SOTS test sets,which is significantly improved compared with most current mainstream algorithms.Moreover,the fog removal effect is good in both the synthetic fog image and the real haze image,and the details of the restored image are clearer and more in line with human visual perception.