SP-IGAN: An Improved GAN Framework for Effective Utilization of Semantic Priors in Real-World Image Super-Resolution

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Generative Adversarial Networks (GANs) show promise in single-image super-resolution (SISR).
• Existing GAN-based SISR methods struggle with locally consistent textures due to limited semantic understanding.
• Improved contextual information comprehension and high-frequency detail acquisition are crucial for advanced SISR models.

Purpose of the Study:

• To propose a novel Semantic Prior-Improved GAN (SP-IGAN) framework for enhanced SISR.
• To address limitations in texture reconstruction and semantic consistency in current SISR approaches.
• To improve the acquisition of high-frequency details in super-resolved images.

Main Methods:

• The SP-IGAN framework integrates contextual semantic information into the Real-ESRGAN model.
• A Graph Convolutional Channel Attention (GCCA) module enhances pixel associations in the main branch.
• An auxiliary branch uses a pretrained segmentation model and Spatial Feature Transform (SFT) layers to inject semantic information into Residual-in-Residual Dense Blocks (RRDB).
• Wavelet loss is incorporated to capture high-frequency details.

Main Results:

• SP-IGAN demonstrates superior performance over state-of-the-art (SOTA) super-resolution models on public datasets.
• Achieved a 0.55 dB improvement in Peak Signal-to-Noise Ratio (PSNR) for X4 super-resolution.
• Increased Structural Similarity Index (SSIM) by 0.0363 compared to the Real-ESRGAN baseline.

Conclusions:

• The proposed SP-IGAN framework effectively improves single-image super-resolution.
• Integrating semantic priors enhances texture consistency and detail reconstruction.
• SP-IGAN offers a promising direction for developing more robust and accurate super-resolution technologies.