DACNet: Single Image Dehazing Network Based on Attention Mechanism
Yongling Liu, Ping Han, Wanwei Wang · 2023
Aiming at the problem that atmospheric scattering model-based haze removal methods are prone to degradation of haze removal performance due to inaccurate parameter estimation, a single-image haze removal network based on the attention mechanism is proposed for direct restoration of haze-free images. The network consists of encoder and decoder, where the traditional convolution is replaced by combining dilated convolutions with different dilated rates for acquiring and processing haze-related image features. Due to the different weighted information contained in different channel features, an increased attention mechanism is used to enhance the feature representation of important channels through adaptive learning channel features, thereby enabling the encoder to pay more attention to channel information in the haze region. In the encoding and decoding structure, dilated convolution and attention mechanism are used in combination for image dehazing for the first time. By targeting the network to pay attention to the regions with high fog distribution while increasing the sensory field to obtain global information. The method presented in this paper improves the ability of network to process hazy images and brings a new breakthrough in the field of hazy image processing. Experiments show that the algorithm proposed in this paper achieves significant improvements in both subjective and objective indicators comparing with many state-of-the-art methods.