Proximal Mapping-Incorporated Deep Autoencoder Network with Momentum Acceleration Method for Image Denoising

In the field of image processing, autoencoder networks have emerged as potent approaches for image denoising. However, traditional autoencoder networks often struggle with imprecise noise modeling and slower training convergence, which hinder their denoising performance. To address these limitations...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Data Driven Control and Learning Systems Conference (Online) s. 1322 - 1327
Hlavní autoři: Zhao, Yang, Li, Ming, Huang, Meng, Zang, Qiyan, Wang, Lele, Zhang, Jian, Zhang, Heng
Médium: Konferenční příspěvek
Jazyk:angličtina
Vydáno: IEEE 09.05.2025
Témata:
Autoencoders
Convergence
Deep autoencoder network
Feature extraction
Image denoising
Image reconstruction
Learning systems
Momentum acceleration scheme
Noise
Noise reduction
Proximal mapping operator
Sparse approximation
Training
ℓ 1 -norm regularization
ISSN:2767-9861
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:In the field of image processing, autoencoder networks have emerged as potent approaches for image denoising. However, traditional autoencoder networks often struggle with imprecise noise modeling and slower training convergence, which hinder their denoising performance. To address these limitations, we propose a novel image denoising autoencoder network incorporated with the proximal mapping operator and the momentum acceleration technology. It adopts threefold ideas: 1) a deep autoencoder network that fully extracting latent features for precise reconstruction; 2) a proximal mapping operator to handle ℓ 1 -norm regularization problem, leading to a stable and sparse representation of noise; and 3) the momentum acceleration scheme coupled with mini-batch gradient descent algorithm to accelerate training convergence. Experimental results on real images demonstrate that our proposed autoencoder network significantly outperforms existing approaches.
ISSN:2767-9861
DOI:10.1109/DDCLS66240.2025.11065628