Single-image Super-resolution Based on Generative Adversarial Network with Dual Attention Mechanism

Most of the single-image (SI) super-resolution (SR) methods based on convolutional neural network (CNN) focus on the broader or deeper model architecture design, ignoring the research on the correlation of features in the middle layer, which hinders the representative power of CNN. However, most of the classical SISR methods based on generative adversarial network (GAN) focus on improving generators and loss functions, which are difficult to balance visual quality and objective evaluation. In order to solve these problems, SRGAN-DAM-SISR is proposed in this paper, which is also a typical GAN structure. However, it is different from the classical SRGAN and ESRGAN. We have made many designs and innovations, including improving the traditional feature extraction module of the generator and the discriminator's discriminant target in the GAN, removing the batch normalization (BN) layer, using many loss functions for weighting and using some tricks to improve the performance of the network. Experimental results show the improvement and innovation to enhance the visual effect of the output image, the resulting image quality is also improved.