DACESR: Degradation-Aware Conditional Embedding for Real-World Image Super-Resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Multimodal large models excel at image super-resolution using language conditions.
• Their performance degrades with low-quality or corrupted images.
• Existing methods struggle to adapt these models for degraded image restoration.

Purpose of the Study:

• To enhance the performance of multimodal large models on degraded image super-resolution.
• To develop a novel approach for extracting meaningful features from degraded images.
• To improve the balance between fidelity and perceptual quality in restored images.

Main Methods:

• Revisiting the Recognize Anything Model (RAM) for degraded image recognition via text similarity.
• Developing a Real Embedding Extractor (REE) using a degradation selection strategy and contrastive learning.
• Integrating REE's high-level features into a Mamba-based network via a Conditional Feature Modulator (CFM).

Main Results:

• The proposed REE significantly improves recognition performance on degraded image content.
• The Mamba-based network with CFM effectively restores image textures, yielding visually pleasing results.
• Experimental results show REE helps balance fidelity and perceptual quality in super-resolution.

Conclusions:

• The REE is a promising method for improving degraded image super-resolution.
• Multimodal models integrated with REE demonstrate significant potential for real-world applications.
• The Mamba architecture shows great promise for advanced image restoration tasks.