Image smoothing is an important processing operation that highlights low-frequency structural parts of an image and suppresses the noise and high-frequency textures. In the paper, we post an intriguing question of how to combine the paired unsmoothed/smoothed images and meaningful edge information to improve the performance of image smoothing. To this end, we propose a structure-preserving image smoothing network, which consists of a main interpreter (MI) and an edge map extractor (EME). The network is trained via contrastive learning on the extended BSD500 dataset. In addition, an edge-aware total variation loss function is utilized to distinguish between non-edge regions and edge maps via a pre-trained EME module, therefore improving the capability of structure preservation. In order to maintain the consistency in structure and background brightness, the outputs from MI are used as anchors for a ternary loss in 1:1 paired positive and negative samples. Experiments on different datasets show that our network outperforms state-of-the-art image smoothing methods in terms of SSIM and PSNR.
Bibliographical noteFunding Information:
The work described in this paper was supported by Hong Kong Metropolitan University Research Grant (No. RD/2021/09).
© 2023, The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature.
- Contrastive learning
- Edge map extractor
- Image smoothing
- Main interpreter
- Structure preservation