"Zero-Shot" Super-Resolution Based on SinGAN

SinGAN is a single image generation model, which can be applied to image super-resolution. When SinGAN is directly applied to super-resolution reconstruction, the image quality is poor. In this paper, a new network structure based on SinGAN is proposed, which can learn effective information from a single image more fully. SinGAN adds random noise to the generator of each scale, which is easy to bring bad influence in the super-resolution task of the image. Therefore, the generator of this algorithm removes the input of noise. In order to ensure a depth convolutional neural network supervision, effective gradient flow characteristics and the ability to reuse, the remaining structure and the introduction of dense connections SinGAN generator to accelerate the convergence speed. Furthermore, introduce attention mechanism into the network. The network performance is further improved by learning the weights of feature channels. Experimental results show that this method has better performance and higher efficiency in image super-resolution task.