Publication: Neutron image restoration by means of enhanced patch intensity prior and blind deconvolution
Loading...
Date
2024-11-01
Authors
Khair’iah, Yazid@Khalid
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This study proposes an improved blind deconvolution oriented principally for
deblurring neutron radiography images. The main aim of this study is to estimate a fast,
accurate, and high-quality PSF, which is essential for blind deconvolution. As a result, a
novel regularization method known as the enhanced patch-wise intensity (EPI) image
prior is proposed, which capitalizes on high-intensity pixels in non-overlapping patches.
The EPI prior was derived from observing the characteristics of natural outdoor images,
where the number of high-intensity pixels is greater in the bright channel, and the intensity
of these pixels decreases substantially after blurring. This study demonstrates that
enforcing the sparsity of the EPI prior could accelerate the estimation process for the
intermediate latent image iteration and lead to much more accurate PSF estimation. The
performance of the proposed method was evaluated using two benchmark deblurring
datasets that focused on PSF estimation. Evidently, the proposed technique outperformed
established techniques in both PSNR and SSIM, averaging 27.10 dB and 0.85,
respectively. Also, the technique was more competitive than established methods, and it
performed better in terms of the kernel similarity index and peak signal-to-noise ratio,
averaging 0.9 and 18 dB, respectively. The methods and procedures were evaluated using
several neutron images captured at the TRIGA PUSPATI reactor. In this case, a standard
denoising strategy was applied prior to restoration as part of the image preprocessing
procedures.The results demonstrate notable improvements, with the average SNR increasing from 2.12 to 6.71 dB and image information entropy rising from 4.19 to 6.91.
This indicates successful denoising and quality enhancement. As for the restoration, this
method resulted in an average BRISQUE index of 45.9, highlighting the competitiveness
of the proposed deblurring framework in terms of no-reference measures. In summary,
the method achieved the highest average entropy and contrast values, at 6.95 and 1.03,
respectively. Overall, the proposed method delivered competitive results both visually
and quantitatively. In terms of speed, the proposed method is the second fastest, averaging
187 s, while its standard deviation of 112 s is the lowest, indicating the stability and
consistency of the algorithm.