Attentive U-recurrent encoder-decoder network for image dehazing

摘要

Haze removal is an important pre-processing step in many computer vision tasks. Convolutional neural networks, especially the U-shaped networks, have shown to be effective in image dehazing. Nevertheless, these networks have three main limitations. First, the relevant haze information, e.g. concentration of haze, is totally ignored. Second, spatial inconsistency and information dilution usually occur when the networks refine the dehazed results with a coarse-to-fine strategy. Third, the receptive field of the network is not large enough to capture structural information. Motivated by these problems, a new attentive U-recurrent encoder-decoder dehazing network is presented, which consists of an attentive recurrent network and a U-recurrent encoder-decoder network. By assuming that haze layers with different depths can be detected by multiple stages, we use an attentive recurrent network to generate the haze attention map for guiding the U-recurrent encoder-decoder network with the concentration of haze to better estimate the clear image. Meanwhile, the features for dehazing are further enhanced and the dehazing results are refined in the U-recurrent encoder-decoder network. This design not only enables spatial consistency but also reduces information dilution with short recurrent pathways. Furthermore, a novel residual pyramid pooling module is also proposed and used in the U-recurrent encoder-decoder network, which provides the network with structural information and with an enlarged receptive field. The experimental results demonstrate that our method outperforms state-of-the-art dehazing algorithms on both synthetic and real hazy images.Haze removal is an important pre-processing step in many computer vision tasks. Convolutional neural networks, especially the U-shaped networks, have shown to be effective in image dehazing. Nevertheless, these networks have three main limitations. First, the relevant haze information, e.g. concentration of haze, is totally ignored. Second, spatial inconsistency and information dilution usually occur when the networks refine the dehazed results with a coarse-to-fine strategy. Third, the receptive field of the network is not large enough to capture structural information. Motivated by these problems, a new attentive U-recurrent encoder-decoder dehazing network is presented, which consists of an attentive recurrent network and a U-recurrent encoder-decoder network. By assuming that haze layers with different depths can be detected by multiple stages, we use an attentive recurrent network to generate the haze attention map for guiding the U-recurrent encoder-decoder network with the concentration of haze to better estimate the clear image. Meanwhile, the features for dehazing are further enhanced and the dehazing results are refined in the U-recurrent encoder-decoder network. This design not only enables spatial consistency but also reduces information dilution with short recurrent pathways. Furthermore, a novel residual pyramid pooling module is also proposed and used in the U-recurrent encoder-decoder network, which provides the network with structural information and with an enlarged receptive field. The experimental results demonstrate that our method outperforms state-of-the-art dehazing algorithms on both synthetic and real hazy images.(c) 2021 Elsevier B.V. All rights reserved.