Image Super-Resolution via Dual-Level Recurrent Residual Networks

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep learning feedforward networks for super-resolution map low-resolution (LR) to high-resolution (HR) images.
• Existing feedforward methods struggle to fully capture the interdependence between LR and HR image representations.
• A need exists for super-resolution models that better utilize multi-level image information.

Purpose of the Study:

• To propose a novel dual-level recurrent residual network (DLRRN) for enhanced image super-resolution.
• To improve the generation of high-resolution (HR) images with richer details and superior visual quality.
• To address limitations in current feedforward super-resolution architectures.

Main Methods:

• Introduced residuals into a dual-level architecture, creating the Dual-Level Recurrent Residual Block (DLRRB).
• DLRRB processes information in both LR and HR spaces, utilizing circular signals for mutual guidance.
• Implemented a recurrent mechanism where current layer information is derived from previous layer outputs in both directions (LR to HR, HR to LR).

Main Results:

• The DLRRN demonstrated strong early reconstruction capabilities, progressively restoring high-resolution (HR) details.
• Extensive evaluations on benchmark datasets confirmed the network's effectiveness.
• Achieved superior results in terms of network parameters, visual fidelity, and objective performance metrics compared to existing methods.

Conclusions:

• The proposed DLRRN effectively generates high-resolution (HR) images with enhanced details and visual appeal.
• The dual-level recurrent residual block design successfully leverages multi-scale spatial information.
• DLRRN offers a promising advancement in deep learning-based image super-resolution technology.