Motion deblurring via multiscale residual convolutional dictionary learning

Sparsity-based methods have been extensively used for restoration of blurry images. To overcome the limitation found in patch-based sparse representation models that the shift-invariant properties of images are often ignored, convolutional sparse coding (CSC) models have been developed. However, exi...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Digital signal processing Ročník 165; s. 105337
Hlavní autoři: Zhang, Yi, Chandler, Damon M., Farias, Mylène C.Q.
Médium: Journal Article
Jazyk:angličtina
Vydáno: Elsevier Inc 01.10.2025
Témata:
Convolutional sparse coding
Learned iterative shrinkage/thresholding
Motion deblurring
Residual learning
Residual learning
Learned iterative shrinkage/thresholding
Convolutional sparse coding
Motion deblurring
ISSN:1051-2004
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Sparsity-based methods have been extensively used for restoration of blurry images. To overcome the limitation found in patch-based sparse representation models that the shift-invariant properties of images are often ignored, convolutional sparse coding (CSC) models have been developed. However, existing CSC models often perform less effectively than leading convolutional neural network (CNN) or Transformer models on the image restoration task. We hypothesize that the reasons are two-fold: One reason is that most existing CSC models often represent images via a single scale by using a universal dictionary, while many CNN models extract multiscale features via different convolutional kernels. The second reason is that existing CSC-based models focus on reconstructing images, whereas CNN/Transformer-based networks focus on reconstructing residual features. These findings inspired us to develop a multiscale residual CSC network. Specifically, we employ a U-shape Transformer network for multiscale feature extraction, and we solve the joint residual-CSC optimization problem by using the efficient learned iterative shrinkage/thresholding algorithm. Experimental results tested on various motion-blur image datasets demonstrate the effectiveness of our model as compared with other state-of-the-art deblurring methods. •An advanced multiscale residual convolutional dictionary learning technique was proposed for motion deblurring.•A new CSC model was proposed to reconstruct the multiscale residual features via multi-dimensional dictionaries.•A lightweight deblurring network was designed based on solving the optimization problem of the RCSC model via LISTA.•MRCDL achieves better performance on both synthetic and real motion-blur images as compared to other methods.
ISSN:1051-2004
DOI:10.1016/j.dsp.2025.105337