DeflickerCycleGAN: Learning to Detect and Remove Flickers in a Single Image.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Rolling shutter cameras, common in CMOS sensors, can produce flickering in images due to asynchronous exposure.
• Artificial lighting and power grid fluctuations exacerbate flickering, creating artifacts in digital images.
• Single image deflickering is challenging, especially without prior camera information or paired datasets.

Purpose of the Study:

• To develop an unsupervised framework for end-to-end single image deflickering.
• To address the limitations of existing methods by working with unpaired images.
• To introduce novel loss functions for preserving image quality during flicker removal.

Main Methods:

• Proposed DeflickerCycleGAN, an unsupervised framework utilizing cycle-consistency loss on unpaired images.
• Introduced novel gradient loss and flicker loss functions to prevent edge blurring and color distortion.
• Developed a flicker detection strategy using an ensemble of Markovian discriminators, without additional training.

Main Results:

• DeflickerCycleGAN demonstrated excellent performance in removing flicker artifacts from single images.
• The framework achieved high accuracy and competitive generalization in flicker detection.
• Experimental results on synthetic and real datasets validated the effectiveness of the proposed approach.

Conclusions:

• DeflickerCycleGAN offers an effective unsupervised solution for single image deflickering.
• The method successfully mitigates flicker artifacts while preserving image content and quality.
• The integrated flicker detection strategy provides a robust and efficient means of identifying problematic images.