NAFBET: Bokeh Effect Transformation With Parameter Analysis Block Based on NAFNet

Xiangyu Kong, Fan Wang, Dafeng Zhang, Jinlong Wu, Zikun Liu; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2023, pp. 1603-1612

Abstract


Bokeh effect transformation(BET) aims to transform the bokeh effect of one lens to another lens without harming the sharp foreground regions in the image. Recent studies have shown remarkable success in bokeh effect rendering. However, unlike the traditional bokeh effect rendering task, the BET task needs to transform the image into the bokeh effect of the specified lens. The existing bokeh rendering method is invalid or inefficient for BET, because each pair of lens needs to independently build different model. To address this limitation, we propose NAFBET, a scalable approach than can perform bokeh rendering for multiple lens using only a single model. NAFBET is based on the structure of the image restoration model NAFNet and expands it by adding the source and target parameter analysis block(PAB) to adapt to the BET task. This block can be very convenient to apply in UNet-based model, which can greatly improve BET performance. We did a lot of experiments to prove the effectiveness of our method. In particular, NAFBET won the 1st place in the NTIRE 2023 Bokeh effect transformation Challenge.

Related Material


[pdf]
[bibtex]
@InProceedings{Kong_2023_CVPR, author = {Kong, Xiangyu and Wang, Fan and Zhang, Dafeng and Wu, Jinlong and Liu, Zikun}, title = {NAFBET: Bokeh Effect Transformation With Parameter Analysis Block Based on NAFNet}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops}, month = {June}, year = {2023}, pages = {1603-1612} }