面向动态场景去模糊的对偶学习生成对抗网络

针对动态场景下的图像去模糊问题,提出一种对偶学习生成对抗网络(dual learning generative adversarial network, DLGAN),该网络可以在对偶学习的训练模式下使用非成对的模糊图像和清晰图像进行图像去模糊计算,不再要求训练图像集合必须由模糊图像与其对应的清晰图像成对组合而成. DLGAN利用去模糊任务与重模糊任务之间的对偶性建立反馈信号,并使用这个信号约束去模糊任务和重模糊任务从两个不同的方向互相学习和更新,直到收敛.实验结果表明,在结构相似度和可视化评估方面, DLGAN与9种使用成对数据集训练的图像去模糊方法相比具有更好的性能.

Dual learning generative adversarial network for dynamic scene deblurring

For the problem of dynamic scene deblurring, a dual learning generative adversarial network(DLGAN) is proposed in this paper. The network can use unpaired blurry and sharp images to perform image deblurring calculations in the training mode of dual learning, which no longer requires the training image set to be a pair of blurry and their corresponding sharp images. The DLGAN uses the duality between the deblurring task and the reblurring task to establish a feedback signal, and uses this signal to constrain the deblurring task and the reblurring task to learn and update each other from two different directions until convergence. Experimental results show that the DLGAN has a better performance compared to nine image deblurring methods trained with paired datasets in structural similarity and visualization evaluation.