TDPN: Texture and Detail-Preserving Network for Single Image Super-Resolution.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep convolutional neural networks (CNNs) lead state-of-the-art in single image super-resolution (SISR).
• Existing SISR models often prioritize peak signal-to-noise ratio (PSNR) over perceptual quality, resulting in images lacking texture and detail.
• This leads to perceptually unpleasant recovered images.

Purpose of the Study:

• To develop a Texture and Detail-Preserving Network (TDPN) for SISR.
• To improve the perceptual quality of super-resolved images by focusing on texture and detail preservation.
• To introduce a novel hybrid loss function for enhanced boundary recovery.

Main Methods:

• A two-branch network architecture: one for multi-reception field learning and another for texture/detail learning.
• A texture and detail-learning branch supervised by decomposed ground-truth textures and details.
• A novel hybrid loss incorporating gradient loss to strengthen boundary information and mitigate overly smooth results from Mean Absolute Error (MAE) loss.

Main Results:

• The proposed TDPN significantly enhances perceptual quality in super-resolved images.
• Experimental results on public datasets show superiority over state-of-the-art SISR methods in PSNR, SSIM, and perceptual quality.
• The TDPN method is model-agnostic and applicable to various existing SISR networks.

Conclusions:

• The TDPN effectively addresses the limitations of current SISR models by preserving textures and details.
• The novel hybrid loss function improves boundary recovery, leading to more visually pleasing results.
• The proposed approach offers a versatile solution for enhancing SISR performance across different networks.