CDJNet: Real-world infrared image defocus deblurring via Content-Detail Joint extraction

•CDJNet: A novel CNN-based method for infrared image defocus deblurring.•The Content-Detail Joint Extraction module for effective feature extraction.•The Multi-Scale Feature Aggregation module designed for multi-scale blur kernels.•A dataset for real-world infrared image defocus deblurring. Due to i...

Full description

Saved in:
Bibliographic Details
Published in:Infrared physics & technology Vol. 150; p. 105975
Main Authors: Zhang, Zhuo, Liu, Chuanming, Zhang, Chen, Zhang, Haoran, Hong, Wenqing
Format: Journal Article
Language:English
Published: Elsevier B.V 01.11.2025
Subjects:
Defocus deblurring
Image restoration
Infrared imaging
Neural networks
Image restoration
Neural networks
Defocus deblurring
Infrared imaging
ISSN:1350-4495
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:•CDJNet: A novel CNN-based method for infrared image defocus deblurring.•The Content-Detail Joint Extraction module for effective feature extraction.•The Multi-Scale Feature Aggregation module designed for multi-scale blur kernels.•A dataset for real-world infrared image defocus deblurring. Due to improper focusing or variations in depth of field, defocus blur frequently occurs in infrared images, making defocus deblurring crucial for improving infrared image quality. While recent deep learning-based methods have shown promising results in defocus deblurring for visible light, their application in the real-world infrared domain is limited by factors such as insufficient textural details, spatially variant blur kernels, and scarce real-world defocused infrared data. To address these issues, we propose CDJNet, a novel convolutional neural network designed for real-world infrared defocus deblurring. To yield more informative features for promising deblurring, the network features the CDJE module that captures high- and low-frequency information using two branches of convolutions: differential convolution in the high-frequency branch for detail preservation and large separable kernel convolution in the low-frequency branch for contextual understanding. Additionally, the proposed MS-FA module incorporates multi-scale pyramid sampling operations to achieve comprehensive aggregation of hierarchical features, enabling effective handling of variable-sized blur kernels. Furthermore, we construct the SIIDD dataset, which contains real-world defocused-clear infrared image pairs, to enhance the network’s performance in real-world scenarios. Compared to synthetic defocused images, our method achieves a 23% improvement in PSNR and a 17% improvement in SSIM. When evaluated on real-world defocused images, it shows enhancements of 32% in LPIPS, 21% in NIQE, and 95% in Mean Gradient. These results demonstrate that our method is effective for defocus deblurring in infrared images.
ISSN:1350-4495
DOI:10.1016/j.infrared.2025.105975