Advancing Real-World Stereoscopic Image Super-Resolution via Vision-Language Model

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Vision-language models have shown success in computer vision.
• Exploiting semantic language knowledge for stereoscopic image super-resolution (SR) is challenging.

Purpose of the Study:

• To propose a vision-language model-based stereoscopic image super-resolution (VLM-SSR) method.
• To leverage semantic knowledge from CLIP for training-free stereoscopic image SR.

Main Methods:

• Utilized CLIP's semantic knowledge via visual prompts for region similarity inference.
• Developed a prompt-guided information aggregation mechanism for inter-view information capture.
• Implemented a cognition prior-driven iterative enhancement for optimizing fuzzy regions.

Main Results:

• The VLM-SSR method effectively enhances stereoscopic images.
• Experimental results on four datasets validate the proposed approach.
• Achieved improved stereoscopic image super-resolution using semantic language knowledge.

Conclusions:

• The proposed VLM-SSR method successfully integrates vision-language models for stereoscopic image enhancement.
• The training-free approach offers a novel way to exploit semantic knowledge in SR tasks.
• Demonstrated the effectiveness of prompt-guided and cognition-driven mechanisms for stereoscopic SR.