Training ESRGAN with multi-scale attention U-Net discriminator.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Super-resolution (SR) aims to reconstruct high-resolution images from low-resolution inputs.
• Existing SR models often struggle with perceptual quality and realistic texture generation.
• Accurate differentiation between image subjects and backgrounds is crucial for effective SR.

Purpose of the Study:

• To introduce MSA-ESRGAN, a novel super-resolution model enhancing perceptual image quality.
• To leverage a multi-scale attention U-Net discriminator for improved subject-background differentiation.
• To establish a new benchmark for blind super-resolution tasks.

Main Methods:

• Developed MSA-ESRGAN, integrating a multi-scale attention U-Net discriminator.
• Initialized the generator with a pre-trained Real-ESRNET model for fair comparison.
• Trained the model on the DIV2K dataset using the Adam optimizer and exponential moving average (EMA).

Main Results:

• MSA-ESRGAN surpassed traditional and state-of-the-art SR models in NIQE, PSNR, and SSIM scores.
• Evaluations on BSD100, Set5, Set14, Urban100, and OST300 datasets confirmed superior performance.
• Subjective evaluations validated enhanced visual quality, texture preservation, and realism.

Conclusions:

• MSA-ESRGAN effectively enhances perceptual image quality and realism.
• The multi-scale attention U-Net discriminator is critical for the model's superior performance.
• MSA-ESRGAN provides a robust benchmark for blind super-resolution, outperforming existing methods.