Super-Interpolation With Edge-Orientation-Based Mapping Kernels for Low Complex 2× Upscaling.

Area of Science:

• Computer Vision
• Image Processing
• Digital Signal Processing

Background:

• Ultra-high-definition (UHD) video services necessitate super-resolution (SR) techniques to generate high-resolution (HR) images from low-resolution (LR) inputs.
• Existing SR methods often require significant computational resources, posing challenges for limited hardware and software environments.

Purpose of the Study:

• To introduce a novel, fast, and low-complexity SR method named super-interpolation (SI) suitable for UHD video generation on resource-constrained devices.
• To unify interpolation and quality enhancement into a single SR process.

Main Methods:

• The SI method employs edge-orientation (EO)-based pre-learned kernels, combining interpolation simplicity with SR quality enhancement.
• It involves an offline training phase where LR image patches are clustered by EO, and class-dependent linear mapping functions are learned.
• During the online up-scaling phase, an appropriate learned function is applied to LR patches based on their EO to generate HR patches directly.

Main Results:

• The SI method demonstrated the smallest running time and required relatively small hardware resources compared to ten state-of-the-art SR methods.
• It outperformed six methods in average Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM).
• The method showed competitive or slightly lower PSNR/SSIM performance compared to the remaining methods.

Conclusions:

• The proposed SI method provides an efficient solution for real-time SR tasks in UHD video processing.
• It offers a practical approach for achieving high-quality image up-scaling with reduced computational complexity and hardware demands.