Domain-connected gradient-enhanced transformer network for robust image denoising

Aiming at the problem that the traditional Transformer-based image denoising methods tend to ignore the redundancy of feature information in the encoding and decoding stages, and cannot well combine prior to improve the efficiency of noise feature extraction, we propose a new Transformer denoising network based on connectivity domain gradient extraction (Connectivity Domain Gradient Extraction Module Enhanced Transformer). The denoising network mainly consists of a distributed adaptive module, a connectivity domain gradient extraction module, and a gate feed-forward and multi-scale unit-based Transformer. The distributed adaptive module is used for multi-scale feature extraction from the input image. As the input layer to the Transformer, it provides rich feature representations for the subsequent network. The connectivity domain gradient extraction module performs averaging on the extracted gradient information during the model encoding stage, effectively reducing the redundancy of feature information transferred by skip connections. Moreover, by introducing gradient prior information, it further improves the network’s feature extraction efficiency. The gate feed-forward and multi-scale unit-based Transformer utilizes the gating mechanism to improve the feed-forward network of the Transformer, remarkably enhancing the filtering ability of image features and strengthening the learning ability of the entire network for noise features. The multi-scale unit ensures that the extracted features simultaneously possess both global and local characteristics. Extensive experiments on four synthetic and three real noise datasets verify the effectiveness. Compared with other advanced networks, our network has a higher peak signal-to-noise ratio, structural similarity index measure, and lower learned perceptual image patch similarity, indicating excellent feature filtering ability. It cannot only maintain the perceptual quality of images but also exhibit better performance at the pixel level and in terms of structure.