Wavelet-Based Dual Recursive Network for Image Super-Resolution.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Single-image super-resolution (SISR) has advanced, but deep learning methods face computational challenges for real-world applications, especially on mobile devices.
• Existing SISR methods often require significant computational resources, limiting their practical deployment.

Purpose of the Study:

• To develop an efficient and time-saving SISR approach with fewer parameters and faster inference.
• To enable practical super-resolution applications on resource-constrained devices like mobile phones.

Main Methods:

• A novel wavelet transform-based network architecture is proposed, performing image super-resolution (SR) in the wavelet domain.
• The low-resolution (LR) image is decomposed into wavelet coefficients (WCs), and the network predicts high-resolution (HR) WCs for reconstruction.
• Introduced two novel modules: wavelet feature mapping block (WFMB) and wavelet coefficients reconstruction block (WCRB), along with a dual recursive framework for joint learning.

Main Results:

• The proposed method achieves efficient and accurate reconstruction of HR WCs.
• Experimental results demonstrate superior performance compared to state-of-the-art methods.
• Achieved over a 2x reduction in model parameters and computational complexity.

Conclusions:

• The wavelet transform-based network offers a computationally efficient solution for SISR.
• The novel WFMB and WCRB modules, combined with the dual recursive framework, effectively enhance WC prediction.
• This approach makes advanced SISR more accessible for real-world applications, particularly on mobile platforms.