Image Deblurring Based on Generative Adversarial Networks

Abstract

Image deblurring technology uses deep learning method to solve the blurry problem of single image , which is a challenging problem in the field of computer vision. In recent years, the rapid development of deep learning and computer vision has promoted the performance of blur processing algorithm. From the perspective of deep learning, the article studies on the image deblurring problem, and uses convolution neural network to achieve the purpose of image deblurring. Aiming at the problem that the scale of single deblurring using multi-scale network is huge, and the important feature information is not fully used, this paper proposes a deblurring algorithm based on generative adversarial networks. The model uses feature pyramid network as a framework instead of the multi-scale input, which effectively reduces the size of network and accelerates the training speed. In order to make better use of feature information, the attention mechanism and dual scale discriminator are introduced into the network. In order to make the training process more stable, the algorithm improves the discriminator loss, using the least squares and relativistic combination. The experimental results show that the image deblurring algorithm based on the generative adversarial network achieves better restoration effect than other algorithms.