DAFNet: A novel image restoration model with mixed SimpleGate

As is known, image restoration is a classical research area in computer vision. Recently, high-performance image restoration algorithms have often been derived from improvements to the Transformer architecture. However, since the Transformer is not designed for processing images, the self-attention mechanism and FFN have flaws when processing images. Recent studies have found that the SimpleGate activation function outperforms other activation functions in image restoration tasks. However, SimpleGate's shortcomings in linear feature mapping limit network generalization ability and affect performance. Due to this, we propose DAFNet image restoration network, utilizing the improved NormSimpleGate (NSG) activation function and features designed in a network structure called DAFBlock tailored for NSG. DAFBlock has improved the self-attention mechanism and FFN for images. DAFNet is evaluated on public datasets for image denoising, deblurring, and deraining tasks, and experimental results show that DAFNet's PSNR metric outperforms all other compared algorithms, with improvements of 0.04 dB, 0.22 dB, and 0.18 dB over the second-best results.