The key contribution of the paper is...
1.
Attention mechanism that can efficiently represent the contributions of neighbor frames for reconstuction.
2.
Residual map to handle misalignments
The introduced network architecture based on U-net with numerous skip connections. The network gets input of consecutive frames of noisy image and an estimated noise map. The novel part is that when transpose convolution, there are attention module for each upsampling stage where it learns the attention of consecutive frames and applies it. Attention module contains two types of attention, each for temporal and spatial correlation.
The paper shows results on both synthetic and real-world noisy data and shown that AME-KPN outperforms KPN and MKPN. Also through ablation studies, the paper shows that each attention and the residual map all contribute to performance.
