Maritime defogging is critical for navigation safety, environmental monitoring, and smuggling surveillance. However, existing methods often struggle with the unique challenges posed by marine environments, such as low contrast, uneven fog density, and dynamic illumination changes. To solve these issues, we propose the sparse semantic feature-guided transformer (SSFT), a novel framework that integrates multi-scale feature representation, deformable convolution, and sparse attention module for maritime defogging to: (a) learn multi-scale representation, (b) improve computation efficiency, and (c) achieve adaptive feature extraction. Extensive experiments conducted on benchmark datasets, such as maritime pods image dataset (MPID), RESIDE, and ship object detection (SOD) dataset, demonstrate the superior performance of SSFT over existing mainstream methods. Source code and pre-trained models are available at https://github.com/he13689/WSDETR.
Li et al. (Wed,) studied this question.