Total variation constraint GAN for dynamic scene deblurring

Abstract

Recovering the sharp image solely from the blurry image in dynamic scene is challenging due to the ill-defined nature of the problem. Through Wasserstein distance and L1 norm of total variation combined regularization, we propose a novel TV-DRGAN optimization framework to obtain a latent sharp image from some observed blurry images. Our method benefits from two aspects: one is the improved object total variation energy to constrain the blurry image, and the other is the generator model combining (UPR)-Blocks and D-Blocks. An (UPR)-Block is composed of one upsampling layer and 3 convolution layers. Consisting of an average-pooling layer and multiple convolution layers, a D-Block comes with different kernel sizes that capture global, and local spatial information of the raw image, separately. By analyzing the information of gradient, we obtain a TV-based based on minimum lower bound of loss function of the generator. Our experiments show that the proposed method outperforms the state-of-the-art conventional algorithms significantly.