Structure-guided multiscale feature interaction based real-time dehazing network

Addressing the current limitations of outdoor dehazing methods, which overlook the specificity of target structures in image content, resulting in incomplete reconstruction of image structures, and the challenge of meeting real-time requirements for subsequent visual tasks with numerous structurally complex dehazing models. In this paper, we propose a structure-guided detail enhancement dehazing module based on prior knowledge of target structure in the image. Simultaneously, we effectively avoid redundant feature information by employing partial convolution, reducing the overall computational complexity of the model. During the downsampling process, we extract multi-scale feature information to enhance the model's feature extraction capability. In the reconstruction stage, we strengthen the interaction between features of different resolutions, achieving an efficient and straightforward real-time dehazing network. Extensive experiments on various dehazing task datasets demonstrate the effectiveness of our proposed model. With only 3.11 million parameters and 13.07 billion FLOPs, our model achieves an SSIM of over 0.988 on the SOTS-outdoor dataset. Larger model structures result in improved performance, as detailed in the experimental section.