A novel image enhancement method based on image decomposition and deep neural networks
•We propose a novel image decomposition-based optimization model for low-light image enhancement, by using the total variation and multi-scale convolutional sparse coding to precisely represent illumination and reflectance layers.•We integrate deep unfolding networks into framework of the proposed o...
Uloženo v:
| Vydáno v: | Pattern recognition Ročník 172; s. 112371 |
|---|---|
| Hlavní autoři: | , |
| Médium: | Journal Article |
| Jazyk: | angličtina |
| Vydáno: |
Elsevier Ltd
01.04.2026
|
| Témata: | |
| ISSN: | 0031-3203 |
| On-line přístup: | Získat plný text |
| Tagy: |
Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
|
| Shrnutí: | •We propose a novel image decomposition-based optimization model for low-light image enhancement, by using the total variation and multi-scale convolutional sparse coding to precisely represent illumination and reflectance layers.•We integrate deep unfolding networks into framework of the proposed optimization algorithm.•The deep unfolding network automatically estimates the priors from training samples.•Experimental results demonstrate that the proposed method outperforms state-of-the-art image enhancement methods in terms of visual quality and robutsness.
Image decomposition and deep learning are active research areas in computer vision tasks, such as cartoon texture decomposition, low-light image enhancement, rain streak removal, image recovery, etc. This paper proposes a novel low-light image enhancement method by joining image decomposition and deep neural network techniques. We introduce a new image decomposition-based optimization model by incorporating the Tikhonov regularization and multi-scale convolutional sparse coding (MSCSC) to enhance image visual effects. To enhance robustness performance, we introduce a noise-free image decomposition error term to effectively suppress noise in low-light images. To effectively implement the proposed method, we incorporate a deep-unfolding neural network and an adaptive denoiser into the alternating direction method of multipliers (ADMM) framework. Since the deep unfolding network can effectively simulate the optimization algorithm process, the interpretability of the network model is increased. Moreover, through end-to-end training, we can automatically estimate the two priors and parameter settings from training samples. Finally, qualitative and quantitative experiments demonstrate that the proposed method outperforms state-of-the-art image enhancement methods in terms of visual quality and robustness. The source code is available at https://github.com/cassiopeia-yxx/LLIE. |
|---|---|
| ISSN: | 0031-3203 |
| DOI: | 10.1016/j.patcog.2025.112371 |