Boosting attention fusion generative adversarial network for image denoising

Boosting has received considerable attention to improve the overall performance of model in multiple tasks by cascading many steerable sub-modules. In this paper, a boosting attention fusion generative adversarial network (BAF-GAN) was proposed, which allows boosting idea and attention mechanism modeling for high-quality image denoising. Specifically, several boosting module groups (BMGs) with group skip connection were employed to form denoiser. Each BMG contains some boosting attention fusion blocks (BAFBs). Each BAFB consists of parallel spatial attention unit and channel attention unit interleaved connection. Moreover, the multi-dimensional inner skip connection within BAFB can carry abundant informative features. Besides, spatial and channel attention mechanisms were also embedded in the discriminator to enhance its ability of discriminating various dimensional information. Meanwhile, a new loss function was given to assist the training process of the model. BAF-GAN can be applied to remove image noise, e.g., Gaussian noise and mixed noise. Comprehensive experiment results demonstrate that the BAF-GAN has the state-of-the-art performance.