GUFORMER: a gradient-aware U-shaped transformer neural network for real image denoising

Real-world image denoising is important in image analysis such as image enhancement, object tracking, and recognition. However, the traditional deep learning-based methods suffer from insufficient ability to perceive global information and long inference time. Aiming at this issue, we construct a Transformer-based U-shaped denoising network (GUFORMER) and implement its deployment and inference acceleration on edge computing platforms. Specifically, the distributional adaptive module is introduced to promote the network’s ability to adapt to the complex distributional characteristics of real image noise. In addition, the improved Swin Transformer is utilized to enhance the perception of global feature information. Meanwhile, the limitation of Transformer in extracting local feature information is compensated by combining with multi-scale attention unit. Furthermore, the skip connection of the U-Net is optimized by the connected domain gradient extraction module to extract fine-grained features passed from the encoder to the decoder, thus reducing the redundancy of feature information and improving the network’s feature extraction capability. The proposed method has been verified on the real-world image datasets such as PolyU, SIDD, DND, and RNI15. Extensive experimental results show that the proposed method can effectively preserve image details while maintaining a higher inference speed.