Image deconvolution with multi-stage convex relaxation and its perceptual evaluation

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Image deconvolution is often addressed using non-convex regularization, posing challenges for optimization.
• Existing methods may struggle with effective noise removal and artifact control.

Purpose of the Study:

• To propose a new image deconvolution method utilizing multi-stage convex relaxation.
• To introduce a novel perceptual evaluation metric for deconvolution quality.

Main Methods:

• Employed multi-stage convex relaxation, a machine learning technique, to solve non-convex regularization problems iteratively.
• Developed the transduced contrast-to-distortion ratio (TCDR) metric based on a human vision system (HVS) model.

Main Results:

• The proposed deconvolution method demonstrated excellent performance in noise reduction and artifact suppression.
• The TCDR metric proved sensitive to ringing and boundary artifacts and efficient to compute.
• Comprehensive evaluations using VSNR and TCDR confirmed improved visual quality with low distortions.

Conclusions:

• The multi-stage convex relaxation approach offers a robust solution for image deconvolution.
• The TCDR metric provides an effective tool for perceptual assessment of deconvolution outcomes.
• The method significantly enhances the visual quality of deblurred images.