Frequency-domain loss function for deep exposure correction of dark images

Area of Science:

• Computer Vision
• Image Processing

Background:

• Classical denoising filters struggle with thresholding and frequency estimation in low-light images.
• Deep learning models for image translation often fail to explicitly constrain noise, leading to suboptimal results.

Purpose of the Study:

• To develop an improved method for exposure correction of dark, blurry, and noisy images captured in uncontrolled, low-light environments.
• To enhance the quality of images processed by deep learning networks for low-light conditions.

Main Methods:

• Proposed a novel Discrete Cosine Transform (DCT)/Fast Fourier Transform (FFT)-based multi-scale loss function.
• Integrated the proposed loss function with traditional losses to train a network for image translation.
• Ensured the loss function is end-to-end differentiable, scale-agnostic, and applicable to both RAW and JPEG formats.

Main Results:

• The proposed DCT/FFT-based multi-scale loss function significantly improved the training of deep networks for low-light image enhancement.
• Achieved state-of-the-art performance on quantitative metrics and subjective tests for image exposure correction.
• Demonstrated the generic applicability of the loss function across different image formats and existing frameworks.

Conclusions:

• The novel multi-scale loss function effectively addresses the limitations of existing methods for low-light image correction.
• The proposed approach leads to visually pleasing and high-quality outputs by explicitly handling image noise and blur.
• This method offers a versatile and efficient solution for enhancing images captured in challenging low-light conditions.