双视觉注意网络的联合图像去雾和透射率估计

Joint transmission map estimation and image dehazing using dual vision attention network

为充分挖掘和利用透射率估计及图像去雾过程中捕获信息的相关性,提出了双视觉注意网络的联合图像去雾和透射率估计算法。它包括图像去雾层及透射率估计层,且各层均包含循环注意网络及编码解码网络。图像去雾层在透射率图的监督下,由循环注意网络生成雾浓度注意图,并引导后续的编码解码网络估计去雾结果。透射率估计层在真实无雾图像的监督下,由循环注意网络生成场景注意图,并引导后续的编码解码网络估计透射率。在此基础上,进一步利用循环单元实现图像去雾层及透射率估计层的信息交互,以便在估计场景注意图的过程中能利用到雾浓度信息,在预测雾浓度注意图的过程中能学习场景信息。实验表明算法在合成雾图及真实图像上均能取得较好的去雾效果,在视觉评价和客观评价方面均优于存在的去雾算法,单张雾图的处理时间仅为0.043 s。能有效用于图像去雾的工程实践中。

To fully exploit the dependency of information captured from transmission map estimation and image dehazing,this paper proposes a dual vision attention network for jointly estimating transmission map and clear image from hazy image. The network consists of an image dehazing layer and a transmission estimation layer,with each layer containing a recurrent attentive network and an encoder-decoder network.The haze attention map is generated in the image dehazing layer under the supervision of the transmission map in the recurrent attentive network and guides the subsequent encoder-decoder network to estimate the dehazing result. Similarly,with clear image supervision,the recurrent attentive network in the transmission estimation layer generates the scene attention map and guides the subsequent encoder-decoder network to predict the transmission map. Based on this architecture,we adopted the recurrent units in the two recurrent attentive networks to exchange hidden information. The haze concentration information can then be used to estimate the scene attention map,and the scene information can be used to predict the transmission map. The experimental results demonstrate that our method not only achieves a good dehazing effect on both synthetic and real images,but also outperforms the existing methods in terms of quality and quantity. The average processing time for a single hazy image is 0. 043 s. Therefore,our method can be used in the engineering practice of image dehazing.