Uncovering the Over-Smoothing Challenge in Image Super-Resolution: Entropy-Based Quantification and Contrastive...

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Peak Signal-to-Noise Ratio (PSNR)-oriented super-resolution models are widely used but often produce over-smoothed images.
• Previous analyses focused on model architecture or loss functions, neglecting the impact of data properties on image quality.

Purpose of the Study:

• To identify and address the center-oriented optimization (COO) problem in super-resolution models.
• To propose a novel loss function that mitigates over-smoothing and enhances image details.

Main Methods:

• Quantified data uncertainty using entropy and linked it to the COO problem.
• Developed Detail Enhanced Contrastive Loss (DECLoss) leveraging contrastive learning to reduce distribution variance and entropy.
• Evaluated DECLoss on super-resolution benchmarks, including PSNR-oriented and Generative Adversarial Network (GAN)-based methods.

Main Results:

• Demonstrated that increased data entropy exacerbates the COO problem, leading to over-smoothed outputs.
• DECLoss effectively reduces entropy and improves the perceptual quality of PSNR-oriented super-resolution models.
• When combined with GANs (e.g., RaGAN), DECLoss achieved state-of-the-art results on the Urban100 dataset.

Conclusions:

• The COO problem, driven by data uncertainty (entropy), is a key factor in super-resolution model over-smoothing.
• DECLoss offers an effective solution by explicitly addressing COO, enhancing detail recovery.
• The proposed method shows broad applicability and effectiveness across different super-resolution approaches.