A Light-Weight Hybrid Network for Image dehazing
Xinhao Liu, Bin Chen, Shiqian Wu · 2023
Single image dehazing is a challenging task in the field of computer vision and image processing. Although several advanced dehazing algorithms based on convolutional neural networks (CNNs) have achieved promising results, their limited receptive fields leave room for improvement. Transformers, another new neural network architecture, have shown significant performance in high-level vision tasks, but applying them directly for image reconstruction tasks may lead to coarse details in results. To address this issue, traditional CNN blocks and a Transformer module are combined, then a light-weight dehazing model is proposed. In the proposed model, convolution layers and a Transformer module are combined as an encoder, in which the convolution layers extract local features (details) while the Transformer module gains global features (structures). Finally, a decoder consisting of traditional CNN blocks restores the haze-free image by merging features of both types. The proposed method is evaluated on the SOTS outdoor dataset and real hazy images and outperforms existing methods with less parameters.