基于双尺度特征融合的单幅图像去雾网络

基于深度学习的图像去雾方法在合成数据集上表现良好，但在真实场景中应用时存在去雾不彻底、颜色失真等问题。提出一种新的单幅图像去雾网络，该网络包含特征提取、特征融合2个模块。在特征提取模块中，通过残差密集块和具有空间注意机制的特征提取块分别提取图像的局部特征和全局特征。在特征融合模块中，利用通道注意力机制对局部特征图和全局特征图进行通道加权，并通过卷积操作融合加权后的局部特征图与全局特征图。最后，采用门控网络自适应结合3个不同深度的融合特征图，以恢复高质量的去雾图像。实验结果表明，所提网络在室内数据集下的峰值信噪比(PSNR)和结构相似度（SSIM）分别为33.04 dB、0.983，在HAZERD数据集下的PSNR和SSIM分别比GridDehazeNet网络高出1.33 dB和0.041。同时，该网络的模型参数量和浮点运算数分别为0.34M和16.06×109frame/s，具有较低复杂度，对合成图像和真实图像均可取得理想的去雾效果。

Single Image Dehazing Network Based on Double Scale Feature Fusion

Learning-based dehazing methods perform well with a synthetic dataset but have certain problems, such as residual haze and color distortion, in a real scene.To solve these problems, a single-image dehazing network based on double-scale feature fusion is proposed in this paper.The proposed network consists of two parts:feature extraction and multiscale feature fusion.In the feature extraction stage, the local and global features of the input image are extracted through the residual dense and feature extraction blocks, respectively, using a spatial attention mechanism.In the feature fusion stage, the local and global feature maps are channel weighted using the channel attention mechanism and fused through a convolution operation.Finally, the gated network adaptively combines three feature maps from different depths to restore the high-quality haze-free images.The experimental results show that the Peak Signal-to-Noise Ratio (PSNR) of the proposed network under indoor dataset is 33.04 dB, and the Structural Similarity (SSIM) is 0.983.And the PSNR and SSIM are 1.33 dB and 0.041 higher than that of the GirdDehazeNet network under HAZERD dataset, respectively.In addition, there are 0.34M model parameters, and the Floating Point Operations (FLOPs) of the network occur at a rate of 16.06×10⁹ frames/s, it shows that the proposed network achieves ideal dehazing results for both synthetic and real images with low complexity.