Research on Image Super-resolution Reconstruction Algorithm Based on Improved Generative Adversarial Network

Abstract:
Aiming at the problem that existing image super-resolution reconstruction algorithm, the problems of network training difficulties and unclear texture details of the generated image appear too smooth and lack of authenticity. This paper proposes an improved image super-resolution reconstruction algorithm based on the super-resolution algorithm for generating confrontation networks (SRGAN). First, for generation network structure, the original residual block is replaced with a densely connected residual block, the batch normalization layer removed reduce computational complexity. Second, VGG-19 network used as the basic structure of the discriminant network, adopt average pooling instead of the original fully connected layer to prevent overfitting. In the third loss function: the perceptual loss function, the adversarial loss function, and the content loss function are introduced to form the total objective function of the generator to optimize the model. The improved algorithm uses Charbonnier loss as the content loss function to evaluate the similarity between the generated image and the real image, and uses WGAN-GP theory to optimize model's anti-loss to accelerate the convergence. Experimental results shows the super-resolution reconstruction algorithm proposed our paper superior to current representative algorithms in subjective and objective image evaluation indexes, and it can better synthesize the texture details of the image.