Multi-Scale Feature Enhancement Image Dehazing Network
Bin Xie, Bobo Zhou, Wenhao Kong · 2024
The inverse problem of single image dehazing affects the performance of many subsequent related visual tasks. Lately, there has been a significant upswing in the focus on dehazing methods that harness the power of deep neural networks. However, existing dehazing methods still suffer from detail loss, color distortion, artifacts and other problems. The current study introduces a network designed for image dehazing that employs multi-scale feature enhancement. The network consists of three modules: preprocessing, backbone and post-processing. Firstly, the preprocessing module is employed to derive feature inputs across various scales. Subsequently, the backbone module's grid structure is utilized to integrate feature information from diverse scales, and finally, the image of the dehazing effect is generated by the post-processing module, which optimises the result by reducing artefacts in the output image In order to efficiently aggregate the potential features and extract the local and global shared information related to the location in the image, enhanced parallel attention is introduced to enhance the extraction of features that are more useful, to reduce the loss of details and to improve the quality of the dehazing model. Numerous experimental data confirm that the method proposed in this study exhibits good dehazing performance on the publicly available dataset RESIDE and real hazy images.