Image Super-Resolution Through Compressive Sensing-based Recovery.

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Single image super-resolution (SISR) is an ill-posed problem often requiring large datasets and significant computational resources.
• Deep learning methods dominate SISR but are limited by data and processing power constraints.
• An alternative approach is needed for scenarios with limited data and computational capacity.

Purpose of the Study:

• To develop a novel, efficient image resizing method for super-resolution inspired by compressive sensing (CS).
• To address the limitations of data-hungry deep learning models in image super-resolution tasks.
• To provide a high-resolution (HR) image recovery technique without reliance on prior training data.

Main Methods:

• The study frames image super-resolution as a CS recovery problem, treating the low-resolution (LR) image as a compressed measurement.
• A deterministic binary block diagonal (DBBD) measurement matrix is employed to preserve visual similarity between LR and HR images.
• The HR image's sparse representation is recovered using the DBBD matrix and a sparsification matrix, followed by dense HR image reconstruction.

Main Results:

• The proposed CS-inspired method was applied to both medical and non-medical images.
• HR images generated by the proposed method were compared against traditional proximal, bilinear, and bi-cubic interpolation techniques.
• The CS-inspired method demonstrated superior performance in delivering high-resolution images compared to traditional interpolation methods.

Conclusions:

• The novel CS-inspired image super-resolution technique effectively recovers high-resolution images.
• The method's superiority is attributed to the unique use of the DBBD matrix and CS recovery algorithm.
• This approach provides a viable alternative for image super-resolution without requiring extensive training datasets or high computational power.