Celebrating Samsung R&D Institute China-Beijing's Excellent Performance at NTIRE @CVPR 2025

The New Trends in Image Restoration and Enhancement (NTIRE 2025) workshop, held alongside the Computer Vision and Pattern Recognition Conference (CVPR 2025), served as a key venue for sharing the latest advances in image processing and restoration. Samsung R&D Institute China-Beijing (SRC-B) participated in multiple challenges and earned top-tier results across several tracks. This remarkable performance highlights SRC-B’s relentless pursuit of innovation and its growing leadership in the field of image restoration and enhancement.

1. Outstanding Achievements in NTIRE@CVPR 2025

1）Real-World Face Restoration Challenge

SRC-B secured 1^st place by demonstrating exceptional capabilities in blind face restoration. Leveraging a novel divide-and-conquer strategy, the team decomposed the complex task into smaller sub-problems, enabling a more targeted and effective restoration pipeline. This innovative approach led to superior results in both accuracy and perceptual quality.

2）RAW Restoration Challenge: Track 2) Restoration - Efficiency & General

SRC-B achieved 1^st place in both the Efficiency and General tracks, showcasing its strength in balancing performance and optimization. The team introduced the NafBlock architecture from NAFNet, applying a narrow-and-deep design principle that effectively reduced parameters while maintaining high restoration quality. In the general track, SRC-B further refined the model by minimizing channel width and encoder-decoder blocks, enhancing efficiency without sacrificing fidelity.

3）Image Denoising Challenge (Noise Level=50)

SRC-B secured 1^st place by delivering outstanding denoising results under high noise conditions. By integrating the strengths of the transformer-based Restormer and the convolutional NAFNet, the team achieved unparalleled performance, effective setting a new benchmark in the field of image denoising.

4）Image Super-Resolution (x4) Challenge - Restoration Quality

SRC-B earned 1^st place in the Restoration Quality track with a hybrid architecture that combined the transformer-based HAT and convolutional NAFNet. This approach enabled the generation of highly realistic texture details, surpassing existing methods in delivering perceptually rich and accurate super-resolution results.

5）HR Depth from Images of Specular and Transparent Surfaces

SRC-B secured 1^st place by applying structural priors from monocular depth estimation models within a cascaded framework. The team utilized pre-trained Vision Transformer (ViT) encoder features from DepthAnything V2, embedding them into a multi-scale feature pyramid to provide fine-grained per-pixel guidance for stereo matching. To further enhance performance, synthetic stereo samples rendered in Blender using the AI2THOR dataset were incorporated during training.

2．Technical Excellence and Innovation

SRC-B's success in these challenges was driven by a combination of innovative architectural designs and meticulous problem-solving strategies. Key innovations include:

Divide-and-Conquer Strategy: SRC-B tackled complex image restoration tasks by breaking them into smaller, manageable sub-tasks, enabling more comprehensive and efficient solutions.

Hybrid Network Architectures: By integrating transformer-based and convolutional networks, SRC-B achieved state-of-the-art results across denoising, super-resolution, and RAW image restoration.

Efficiency and Fidelity: A focus on parameter efficiency and narrow-and-deep network design allowed SRC-B’s models to maintain high restoration quality while being optimized for real-world deployment.

Cascaded Framework for Depth Estimation: Leveraging pre-trained Vision Transformer (ViT) encoder features and multi-scale feature pyramids, SRC-B provided per-pixel guidance for stereo matching, significantly boosting depth estimation accuracy in specular and transparent surface scenarios.

3．Future Prospects and Commercialization

SRC-B’s groundbreaking achievements hold significant potential for the future of image processing across various real-world applications. The developed technologies are well-positioned for commercialization in areas such as:

Mobile Camera ISP: Enhanced denoising and super-resolution capabilities can substantially improve smartphone camera performance, delivering higher-quality images even in challenging lighting or motion conditions.

RAW Image Processing: Advancements in RAW restoration open new opportunities for professional photography and computational imaging, enabling richer detail and better post-processing flexibility.

Zoom and Beyond: SRC-B’s super-resolution technologies can be applied to improve digital zoom capabilities in consumer cameras, allowing users to capture high-fidelity images even at high magnifications.

Robotics and Scene Understanding: The depth estimation techniques developed by SRC-B enhance robotic perception in complex environments, enabling more accurate navigation, manipulation, and interaction in real-world scenarios.

SRC-B's remarkable performance at NTIRE@CVPR 2025 stands as a testament to the team's dedication, innovation, and technical excellence. By tackling complex challenges and consistently pushing the boundaries of image restoration, SRC-B is helping to set new benchmarks in the field. As these cutting-edge technologies move from research into real-world applications, they hold the potential to transform how we capture and enhance visual information, further reinforcing SRC-B's growing leadership in the industry.