Generative Adversarial Network for Image Deblurring Using Content Constraint Loss

摘要

The research of image deblurring plays an important role in the digital image processing. In order to reduce image blurring problems, a content constraint loss (CCL) function in the generative adversarial network (GAN) is proposed. The SSIM loss and the perceptual loss constitute the CCL function, which makes the trained generative model stable. The CCL function as the content constraint loss component and the adversarial loss component constitute the total loss. The total loss is optimized by the iterative training to further improve the stability of the network model, and the image blurring will be reduced. In the test experiment of the open source image dataset MNIST, CIFAR10/100 and CELEBA, the CCL function is used as the content constraint loss component of the generative adversarial network, the effect of image deblurring has obvious promotion in the structural similarity measure and visual appearance.