Project Icon

CAT

Improve Image Restoration with the Cross Aggregation Transformer Model

Product DescriptionThe Cross Aggregation Transformer enhances image restoration by incorporating innovative features such as Rectangle-Window Self-Attention and the Locality Complementary Module, improving on traditional Transformer models. This approach surpasses standard methods by overcoming the limitations of local square windows through horizontal and vertical rectangle attention and cross-window feature aggregation. The Axial-Shift operation further refines window interaction, supporting long-range dependencies. Furthermore, integrating the CNN's inductive bias into the Transformer allows for a hybrid global-local interaction framework, showcasing superior performance in extensive testing compared to recent state-of-the-art methods.
Project Details