Exploiting Raw Images for Real-Scene Super-Resolution.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Single image super-resolution (SISR) algorithms often fail on real-world data due to limitations with synthetic training datasets.
• Existing methods do not fully leverage the visual information captured by digital cameras, particularly in raw image formats.

Purpose of the Study:

• To bridge the gap between synthetic and real-world data for single image super-resolution.
• To develop a robust SISR algorithm applicable to diverse real-world scenes and camera types.
• To enhance the utilization of camera-captured visual information for improved image restoration.

Main Methods:

• Generation of realistic training data by simulating the digital camera imaging process.
• Development of a two-branch convolutional neural network to process raw image radiance information.
• Introduction of a dense channel-attention block for superior image restoration.
• Implementation of a learning-based guided filter network for accurate color correction.

Main Results:

• The proposed model demonstrates generalization across different camera types without specific training.
• The algorithm effectively recovers fine details and sharp structures in real-scene images.
• Experimental results confirm high-quality performance in real-scene single image super-resolution tasks.

Conclusions:

• The developed method significantly advances real-scene single image super-resolution by addressing data realism and information utilization.
• The approach offers a practical solution for enhancing image quality from digital cameras in real-world scenarios.
• The model's ability to generalize across cameras makes it a versatile tool for various applications.