A Deep Learning-Based Two-Branch Generative Adversarial Network for Image De-Raining.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Raindrops degrade image quality by causing blurriness and distortion due to light scattering and absorption.
• Existing image de-raining methods struggle to effectively restore details and prevent information loss.

Purpose of the Study:

• To propose a novel generative adversarial network (GAN) for robust image de-raining.
• To enhance image quality by mitigating the adverse effects of raindrops on visual data.

Main Methods:

• A novel GAN architecture featuring a dual-branch generative network (A-branch and U-branch) and an adversarial network.
• The A-branch utilizes multi-scale and residual attention modules; the U-branch employs an encoder module to preserve details.
• A relative discriminator with mean squared error loss is used to improve de-raining performance and prevent gradient vanishing.

Main Results:

• The proposed method demonstrates superior performance in visual and quantitative evaluations across three benchmark datasets.
• Achieved average improvements of approximately 5% in Peak Signal-to-Noise Ratio (PSNR), 3% in Structural Similarity Index Measure (SSIM), and 4% in Visual Information Fidelity (VIF) compared to MFAA-GAN.
• De-rained images generated by the proposed method exhibit significantly higher fidelity to original rain-free images.

Conclusions:

• The proposed generative adversarial network effectively removes rain streaks and restores image quality.
• The novel architecture and loss function contribute to state-of-the-art performance in image de-raining tasks.
• This research offers a promising solution for enhancing visual data captured in adverse weather conditions.