Hierarchical integration transformer for marine snow removal

Transformer has demonstrated excellent performance in image synthesis and object removal tasks, especially in capturing global representations. However, Transformer tend to overlook local image details, and their computational complexity increases quadratically with spatial resolution. To address these issues, we propose a hierarchical integrated Transformer for marine snow removal. Specifically, we leverage the Transformer to capture global information in the latent space and hierarchically integrate it into a CNN-based model to remove marine snow particles and restore image details. In the integration process, we introduce a global-local integration module that effectively combines global and local information across multiple levels from Transformer and CNN respectively. We conduct extensive experiments on public datasets of marine snow, including MSRB and Snowy-VAROS, demonstrating the exceptional performance of our method for marine snow removal.