DDUNet: Dense Dense U-Net With Applications in Image Denoising

Fan Jia, Wing Hong Wong, Tieyong Zeng; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021, pp. 354-364

Abstract


The investigation of CNN for image denoising has arrived at a serious bottleneck and it is extremely difficult to design an efficient network for image denoising with better performance and fewer parameters. A nice starting point for this is the cascading U-Nets architecture which has been successfully applied in numerous image-to-image tasks such as image denoising and segmentation. However, the previous related models often focused on the local architecture in each U-Net rather than the connection between U-Nets, which strictly limits their performances. To further improve the connection between U-Nets, we propose a novel cascading U-Nets architecture with multi-scale dense processing, named Dense Dense U-Net (DDUNet). The multi-scale dense processing connects the feature maps in each level cross cascading U-Nets, which has several compelling advantages: they alleviate the vanishing gradient problem, strengthen feature propagation and encourage feature reuse. Furthermore, we develop a series of related important techniques to improve model performance with fewer parameters. Extensive experimental results on both synthetic and real noisy datasets demonstrate that the proposed model achieves outstanding results with fewer parameters. Meanwhile, experimental results show clearly that the proposed DDUNet is good at edge recovery and structure preservation in real noisy image denoising.

Related Material


[pdf]
[bibtex]
@InProceedings{Jia_2021_ICCV, author = {Jia, Fan and Wong, Wing Hong and Zeng, Tieyong}, title = {DDUNet: Dense Dense U-Net With Applications in Image Denoising}, booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops}, month = {October}, year = {2021}, pages = {354-364} }