Low Light Image Enhancement Algorithm Based on Detail Prediction and Attention Mechanism

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Existing low-light image enhancement (LLIE) algorithms often prioritize brightness, neglecting vital image details like edges and textures.
• This leads to information loss, semantic distortion, and degraded image quality in conventional methods.

Purpose of the Study:

• To propose the Deep Enhancement of Low-Light Image (DELLIE) algorithm, a novel framework for low-light image enhancement.
• To focus on the extraction and fusion of image detail features alongside brightness enhancement.

Main Methods:

• DELLIE employs a deep learning framework involving basic enhancement preprocessing and a detail component prediction model.
• It utilizes a decomposition network to separate the V-channel into reflectance and illumination maps, enhancing the reflectance map.
• The algorithm incorporates an improved adaptive loss function for nonlinear constraint of S and H channels and integrates an attention mechanism.

Main Results:

• DELLIE effectively extracts and recovers image detail information while simultaneously improving luminance.
• Experimental results demonstrate superior performance compared to mainstream LLIE algorithms.
• Average optimization of 1.85% in PSNR, 4.00% in SSIM, and 2.43% in NIQE on benchmark datasets.

Conclusions:

• The proposed DELLIE algorithm offers a significant advancement in low-light image enhancement.
• It successfully addresses the limitations of existing methods by preserving image details and semantics.
• DELLIE provides a robust framework for achieving high-quality enhanced low-light images.