Denoising Method Based on Salient Region Recognition for the Spatiotemporal Event Stream.

Area of Science:

• Computer Vision
• Sensor Technology
• Biomimetics

Background:

• Event cameras, or dynamic vision sensors, offer high-speed, event-based data capture.
• Noise in event camera data, arising from light sensitivity and external factors, degrades perception and algorithm performance.
• Traditional image denoising methods are unsuitable for the address-event representation output of event cameras.

Purpose of the Study:

• To develop an effective denoising method for event camera data.
• To address limitations of existing methods in handling various noise types, including those from flickering lights and diffuse reflections.
• To ensure real event preservation during the denoising process.

Main Methods:

• Proposed a novel event stream denoising method utilizing salient region recognition.
• Developed a new evaluation metric to assess denoising efficacy and real event preservation.
• Tested the method against conventional background activity noise and irregular noise sources.

Main Results:

• The proposed method successfully removes background activity noise and irregular noise from diffuse reflections and flickering light sources.
• The method demonstrates minimal loss of real event data.
• The new evaluation metric provides a quantitative assessment of denoising performance.

Conclusions:

• The salient region recognition-based denoising method offers a significant improvement for event camera data processing.
• This approach enhances the reliability and performance of algorithms using event camera streams.
• The developed evaluation metric aids in comparing and advancing event camera denoising techniques.