Learning a Deep Single Image Contrast Enhancer from Multi-Exposure Images.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Digital images often suffer from poor lighting, limited dynamic range, leading to under-/over-exposure and low contrast.
• Existing single image contrast enhancement (SICE) methods, primarily tone curve adjustments, struggle to reveal details due to information limitations in single images.

Purpose of the Study:

• To propose a novel SICE method utilizing convolutional neural networks (CNNs) for enhanced contrast and detail.
• To address the challenge of creating suitable training datasets for end-to-end CNN learning in SICE.

Main Methods:

• A large-scale multi-exposure image dataset was constructed, comprising 589 high-resolution sequences (4,413 images).
• Multiple image fusion and high dynamic range imaging algorithms were used to generate contrast-enhanced reference images.
• Subjective experiments were conducted to select the best quality reference image for each scene.

Main Results:

• A CNN-based SICE enhancer was trained using the constructed dataset.
• Experimental results demonstrated significant advantages of the proposed CNN method over existing SICE techniques.
• The method effectively improves the contrast of under-/over-exposed images.

Conclusions:

• The proposed CNN-based approach offers a superior solution for single image contrast enhancement.
• The curated multi-exposure dataset facilitates effective training of CNNs for SICE tasks.
• This work advances the field of image enhancement by leveraging deep learning for improved image quality.