.png){kind=icon}
.png){kind=icon}
Return
Ghost-free multi-exposure high dynamic range imaging based on feedback network
.png){kind=icon}
PRE
AI
DOI：10.1007/s00371-023-03072-w.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In multi-exposure HDR imaging of dynamic scenes, two challenging problems need to be addressed: one is the removal of ghost artifacts, and the other is the reconstruction of lost information. Researchers usually adopt feed-forward networks to deal with the two problems in the existing deep learning-based multi-exposure HDR imaging methods. However, feedback mechanism plays an important role in human visual system, and feedback networks achieves efficient performance on other imaging tasks. Therefore, we presented a novel deep learning algorithm based on feedback mechanism to achieve ghost-free multi-exposure HDR imaging. To our best knowledge, it is the first time that feedback mechanism is applied to multi-exposure HDR imaging. We first use an attention-guided encoder to adjust the shallow features of the input images to avoid redundant features that interfere with feature fusion. Then, the adjusted features are fused by a feature merging network. It consists of two blocks: a feedback block and a global feature extracting block. In the feedback block, a hidden state is used in a recurrent structure to achieve the feedback mechanism. It enables the feedback block to use high-level features to enhance its feature fusion capability, and the reconfiguration capability of the network gradually improves during iterations. In the global feature extracting block, the extracted overall information increases the receptive field of the network and helps in the recovery of lost details. Extensive experiments show the superior performance of the proposed method in high-quality HDR images reconstruction, and outperforms state-of-the-art methods considered in comparison.
Keywords:
High dynamic range imaging
Deghosting
Feedback network
Multi-exposure imaging
Journal
IF:
2.9
Papers:
4.6K
Citations:
6.5K
Organization
No organization information available