Enhancing Low-Light Light Field Images With a Deep Compensation Unfolding Network

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Low-light conditions severely degrade the quality of light field (LF) images.
• Restoring LF images under low-light conditions is challenging due to noise and illumination variations.
• Existing methods often struggle to preserve the unique geometric information inherent in LF data.

Purpose of the Study:

• To propose a novel and interpretable end-to-end learning framework for low-light LF image restoration.
• To develop a network that mimics inverse imaging problem optimization in a data-driven manner.
• To enhance the quality and preserve the geometric structure of low-light LF images.

Main Methods:

• Introduced the Deep Compensation Unfolding Network (DCUNet), a multi-stage architecture.
• Utilized intermediate enhanced results for illumination map estimation and subsequent unfolding.
• Incorporated a content-associated deep compensation module to mitigate noise and estimation errors.
• Proposed a pseudo-explicit feature interaction module to exploit LF image redundancy.

Main Results:

• DCUNet demonstrated superior performance over state-of-the-art methods on simulated and real datasets.
• Qualitative and quantitative evaluations confirmed the effectiveness of the proposed framework.
• The method significantly better preserves the essential geometric structure of enhanced LF images.

Conclusions:

• DCUNet offers an effective and interpretable solution for low-light LF image restoration.
• The proposed modules successfully address noise, illumination estimation errors, and leverage LF image characteristics.
• The framework provides high-quality enhanced LF images with preserved geometric integrity.