Dynamic Scene Deblurring Using Enhanced Feature Fusion and Multi - Distillation Mechanism.

Qianyi Zhang,Zhixin Zeng,Yiming Lie,Kang Tane,Ji Wang

DOI: https://doi.org/10.1109/ijcnn52387.2021.9533906

2021-01-01

Abstract:Despite the surges of deep learning-based method in dynamic scene deblurring achieves good performance, the challenges still remain a lot: (a) the running speed is far from the requirement of processing; (b) there will be inevitable information loss along with the deepening of network layers, which will further lead to the deterioration of the quality of the restored pictures. To deal with these challenges, we propose a novel learning-based model. In our method, we integrate two mechanisms for the generator based on the Generative Adversarial Nets (GAN). First, we develop the Enhanced Feature Fusion (EFF) mechanism which aims at providing multi-layer feature information to assist the image restoration. We further design Feature Multi-distillation (FMD) mechanism to filter and well fuse the multi-scale feature maps. By integrating the two mechanisms, the high-level feature maps can be progressively refined and the detailed semantic information can be properly utilized. In addition, we use the double-scale discriminator architecture which could enables the network to observe the image from the perspective of local and global respectively and obtain overall information for the whole image restoration. Extensive experimental results on the GOPRO and Kohler datasets show that our method can approach comparably to the state-of-the-arts in terms of accuracy while consuming much less inference time, which demonstrates that our method acquiring a better trade off between image restoration quality and running speed.