从全局到局部:双注意力融合去雾网络

From Global to Local:A Dual-Attention Fusion Dehazing Network

为了处理现有的基于卷积神经网络去雾方法只使用单一的注意力、很难生成细节生动的清晰图像,且容易导致色彩失真的问题,提出了一个全局与局部注意力融合的图像去雾方法,以获得正常清晰度和无色彩失真的去雾图像。首先利用通道注意力将输入的有雾图像在通道维度切分为两部分,一部分送入通道像素注意力通道抽取局部特征,另一部分送入Transformer通道学习全局特征,然后利用像素注意力对两个通道学习的特征进行融合,将上述模块作为基本单元组合为一个多级U型去雾网络,增加残差连接缓解上下采样导致的细节信息丢失,最后在网络底层加入一个Transformer模块学习全局信息。在多个公开可用的去雾图像数据集RESIDE SOTS Indoor、RESIDE SOTS Outdoor上测试所提方法的有效性,结果表明:对比经典的去雾方法,所提网络生成的图像细节更丰富并且色彩失真最少;在RESIDE SOTS Outdoor数据集上,相比经典的FFA-Net,峰值信噪比提高1.16 dB,相比GridDehazeNet,峰值信噪比提高3.68 dB。提出的全局与局部注意力融合方法能有效地去除雾霾,提升图像的对比度与清晰度,设计的多级U型去雾网络和残差连接结构能够缓解细节丢失,提升去雾效果,获得清晰的图像。

In terms of the problem that the existing dehazing methods based on the convolution neural network employs attention only from a single perspective,causing difficulties in generating a clear image with vivid details and the propensity to give rise to color distortion,this paper proposes a global and local attention fusion image dehazing network is proposed,in order to obtain a dehazing image with normal definition and no color distortion.The input haze image is first divided into two parts in the channel dimension by using the channel attention.One part is sent into the channel pixel attention channel to extract local features,and the other part is sent into the Transformer channel to learn global features.Then,the pixel attention is used to fuse the features learned by the two channels,and the above modules are combined as basic units into a multi-level U-shaped dehazing network,residual connection is added to alleviate the loss of detail information caused by upper and lower sampling,and finally,a Transformer module is added at the bottom of the network to learn global information.The effectiveness of the method proposed is tested on several publicly available dehazing image data sets including RESIDE SOTS Indoor and RESIDE SOTS Outdoor.The results show that compared with the classical dehazing method,the image generated by the method proposed is more detailed and has the least color distortion.On the RESIDE SOTS Outdoor data set,the PSNR is 1.16dB higher than that of the classical FFA-Net,and 3.68dB higher than that of the GridDehazeNet.The global and local attention fusion method proposed in this paper can effectively remove haze and improve the contrast and clarity of the image;the designed multi-level U-shaped dehazing network and residual connection structure can alleviate the loss of details and improve the dehazing effect,so that clear images are obtained.