Noise Removal in the Developing Process of Digital Negatives

Area of Science:

• Digital image processing
• Computational photography
• Image sensor technology

Background:

• Modern digital cameras use color filter arrays (CFAs) on image sensors.
• Image preprocessing, particularly noise reduction, is crucial.
• Existing research often overlooks raw image acquisition realities, focusing on sRGB space.

Purpose of the Study:

• Analyze the optimal placement of denoising and demosaicing algorithms within the raw image processing pipeline.
• Investigate the impact of noise reduction on raw sensor data versus processed image data.
• Simulate noise generation based on ISO sensitivity for CMOS sensors.

Main Methods:

• Developed a diverse database of raw digital negative images and their reference versions.
• Reproduced all stages of raw image processing.
• Simulated sensor noise generation, parameterized by ISO sensitivity.
• Evaluated non-local means algorithms (NLM, BM3D) with various CFA interpolation techniques.

Main Results:

• Noise reduction applied directly to raw sensor data yields significant improvements only for images with high noise levels.
• This benefit is most pronounced in challenging lighting conditions that lead to increased image noise.
• The placement of denoising algorithms within the processing pipeline impacts overall image quality.

Conclusions:

• Applying noise reduction early in the raw image processing pipeline is beneficial for severely noisy images.
• The effectiveness of raw data denoising is highly dependent on the noise level and acquisition conditions.
• Optimizing denoising and demosaicing integration is key for high-quality digital image capture.