Deep Perceptual Image Enhancement Network for Exposure Restoration

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Poor illumination significantly degrades image quality, causing noise, color distortion, and lack of sharpness.
• Existing image restoration methods struggle with challenges posed by underexposed or overexposed regions.

Purpose of the Study:

• To present a novel deep convolutional neural network, the deep perceptual image enhancement network (DPIENet), for end-to-end image enhancement.
• To address limitations in image quality caused by poor illumination conditions.

Main Methods:

• Developed a framework to synthesize multiple exposures from a single image for restoration.
• Introduced a novel loss function approximating the human eye's logarithmic response.
• Utilized a deep convolutional neural network architecture (DPIENet).

Main Results:

• DPIENet demonstrated clear advantages over state-of-the-art techniques in extensive simulations.
• Evaluated on benchmark datasets including MIT-Adobe FiveK, Google high dynamic range, DIV2K, and low light images.
• User studies confirmed the effectiveness of the proposed method.

Conclusions:

• DPIENet offers superior image restoration for low-light conditions.
• The proposed method enhances visibility, contrast, and color vividness in degraded images.
• Potential applications include improving camera technology, consumer photography, and intelligent systems like automated driving and surveillance.