Single image superresolution based on gradient profile sharpness.

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Single image superresolution (SISR) aims to reconstruct a high-resolution (HR) image from a low-resolution (LR) input.
• The ill-posed nature of SISR necessitates effective image priors for accurate reconstruction.
• Existing methods often struggle with preserving edge details and overall image quality.

Purpose of the Study:

• To propose a novel SISR algorithm leveraging gradient profile sharpness (GPS) as an image prior.
• To develop a method for statistically analyzing and estimating the GPS transformation relationship across resolutions.
• To generate superior HR images with enhanced visual fidelity and reduced reconstruction errors.

Main Methods:

• Extraction of GPS, an edge sharpness metric, using triangle and Gaussian mixture models for gradient profiles.
• Statistical study and automatic estimation of the GPS transformation relationship between LR and HR images.
• Development of gradient profile transformation models that maintain shape and magnitude sum consistency.
• Generation of HR image gradient fields using transformed profiles as priors in reconstruction.

Main Results:

• The proposed algorithm demonstrates superior performance in subjective visual quality.
• Objective quality metrics show a significant reduction in reconstruction error compared to state-of-the-art methods.
• The algorithm achieves acceptable computation efficiency, making it practical for real-world applications.

Conclusions:

• The GPS-based image prior effectively addresses the under-determined nature of SISR.
• The novel transformation models ensure accurate gradient field generation for HR reconstruction.
• The proposed method offers a robust and efficient solution for single image superresolution.