Wavelet Pyramid Recurrent Structure-Preserving Attention Network for Single Image Super-Resolution.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Current single image super-resolution (SISR) methods using convolutional neural networks (CNNs) often neglect structural context and detail fidelity, leading to distorted reconstructions.
• Incorporating image priors into CNNs is crucial for improving image reconstruction quality.

Purpose of the Study:

• To propose a novel recurrent structure-preserving mechanism for SISR that leverages the multiscale wavelet transform (WT) as an image prior.
• To introduce the Wavelet Pyramid Recurrent Structure-Preserving Attention Network (WRSANet) for enhanced image reconstruction.

Main Methods:

• The proposed WRSANet utilizes a novel structure scale preservation (SSP) architecture and structure scale fusion (SSF) with inverse WT for recursive low-frequency structure restoration.
• Novel low-to-high-frequency information transmission (L2HIT) and detail enhancement (DE) mechanisms are introduced to improve high-frequency detail fidelity.
• A joint loss function balances low- and high-frequency information fusion, with adaptive hyperparameter tuning during training.

Main Results:

• WRSANet demonstrates superior performance and visual quality compared to state-of-the-art methods on both synthetic and real-world datasets.
• The method excels particularly in preserving context structure and reconstructing intricate texture details.

Conclusions:

• The WRSANet effectively addresses limitations in traditional SISR methods by preserving low-frequency structures and enhancing high-frequency details.
• The proposed approach offers significant improvements in detail fidelity and structural integrity for super-resolved images.