Abstract:
In response to the issue of the uneven handling of the focused and defocused border regions in the multi-focus image fusion task, this study proposes a generative adversarial network (GAN) approach with dual discriminators based on the global feature extraction capability of Transformer. The generator completes the multi-focus image fusion task in an end-to-end manner. This approach leverages the Transformer to obtain global dependencies and low-frequency spatial details, and utilizes a cross-domain cross-attention mechanism to facilitate information interaction between the dual branches of the generator model, obtaining redundant and complementary information from the other channel. The model combines global image information for parameter updates during the learning process, overcoming the aforementioned issue while preserving effective information from the focused regions in the multi-focus image as much as possible. Comparative experiments demonstrate the feasibility and competitiveness of the proposed method.