Improving attention mechanisms in transformer architecture in image restoration

We discuss a problem of improving the quality of images obtained under the influence of various kinds of noise and distortion. In this work we solve this problem using transformer neural network models, because they have recently shown high efficiency in computer vision tasks. An attention mechanism of transformer models is investigated and problems associated with the implementation of the existing approaches based on this mechanism are identified. We propose a novel modification of the attention mechanism with the aim of reducing the number of neural network parameters, conducting a comparison of the proposed transformer model with the known ones. Several datasets with natural and generated distortions are considered. For training neural networks, the Edge Loss function is used to preserve the sharpness of images in the process of noise elimination. The influence of the degree of compression of channel information in the proposed attention mechanism on the image restoration quality is investigated. PSNR, SSIM, and FID metrics are used to assess the quality of the restored images and for a comparison with the existing neural network architectures for each of the datasets. It is confirmed that the architecture proposed by the present authors is, at least, not inferior to the known approaches in improving the image quality, while requiring less computing resources. The quality of the improved images is shown to slightly decrease for the naked human eye with an increase in the channel information compression ratio within reasonable limits.