Modeling Dual-Exposure Quad-Bayer Patterns for Joint Denoising and Deblurring.

Area of Science:

• Computer Vision
• Image Processing
• Sensor Technology

Background:

• Image degradation from noise and blur is a persistent challenge in imaging systems.
• Single-image solutions face a noise-reduction vs. motion-blur tradeoff.
• Multi-image solutions struggle with motion-induced misalignment.

Purpose of the Study:

• To propose a physical-model-based image restoration approach using a novel dual-exposure Quad-Bayer sensor.
• To address the inherent limitations of single-image and multi-image restoration methods.
• To develop a method that integrates complementary noise-blur information within a single image.

Main Methods:

• Utilized a novel dual-exposure Quad-Bayer pattern sensor capturing short and long exposure pairs.
• Introduced a Bayer-to-Quad-Bayer (B2QB) synthesis method for training data simulation.
• Designed a hierarchical convolutional neural network (QRNet) for high-quality RGB image recovery.

Main Results:

• Demonstrated superior performance over state-of-the-art deblurring and denoising methods.
• Achieved high-quality image restoration on both synthetic and real-world datasets.
• QRNet effectively integrated complementary noise-blur information using the dual-exposure sensor.

Conclusions:

• The proposed dual-exposure sensor and QRNet offer an effective solution for image restoration.
• This approach overcomes limitations of existing single- and multi-image techniques.
• The publicly available code, model, and datasets facilitate further research in image restoration.