Real Image Denoising With a Locally-Adaptive Bitonic Filter.

Area of Science:

• Image processing
• Computer vision
• Signal processing

Background:

• Image noise is a pervasive challenge in digital imaging.
• Learning-based methods are current standards but have limitations like data dependency and lack of predictability.
• Non-learning-based filters offer alternatives but often lag in performance.

Purpose of the Study:

• To develop a novel, non-learning-based image noise removal filter.
• To improve upon the traditional bitonic filter with local adaptivity for real-world image noise.
• To achieve high noise reduction performance without compromising processing speed.

Main Methods:

• Developed a novel bitonic filter with a locally adaptive domain.
• Incorporated adjustments for effective application to real image sensor noise.
• Evaluated performance against established filters like block-matching 3D (BM3D) and recent non-learning methods.

Main Results:

• The new bitonic filter significantly improves noise reduction performance.
• Outperforms the block-matching 3D filter in high levels of additive white Gaussian noise.
• Surpasses existing non-learning filters on public datasets with real image noise, even outperforming an enhanced BM3D.
• Achieves performance comparable to optimally trained learning-based methods when trained on unrelated data.

Conclusions:

• The novel bitonic filter provides a predictable, explainable, and entirely local solution for image noise removal.
• It demonstrates robust performance in very high noise levels and challenging scenarios where training data is limited or inappropriate.
• This filter offers a viable alternative to learning-based approaches, especially when predictability and explainability are paramount.