Self-attention negative feedback network for real-time image super-resolution

Abstract In the field of real-time image enhancement, image super-resolution (SR) is an important research hotspot. As an image super-resolution method, deep learning can extract more stable and higher level features. However, image super-resolution processing is an ill posed problem. Due to the lack of self-attentional negative feedback mechanism, the existing methods can not better constrain the mapping space from low-resolution image to high-resolution image, so generated high-resolution (HR) image does not conform to human visual perception. Therefore, this paper proposes a self-attention negative feedback network (SRAFBN) for realizing the real-time image SR. The network model constrains the image mapping space and selects the key information of the image through the self-attention negative feedback model, so that higher quality images can be generated to meet human visual perception. Specifically, a Recurrent Neural Network (RNN) is firstly utilized to construct multiple negative feedback modules, and each module generates a HR image. Then, the key information of the generated image is extracted by the self-attention mechanism. Finally, the extracted information is fused to generate the final HR image. Experimental results prove the proposed SRAFBN network model not only has higher PSNR and SSIM, but also can reconstruct more realistic and clear real-time images.