Single-Image Visibility Restoration: A Machine Learning Approach and Its 4K-Capable Hardware Accelerator.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Machine vision algorithms are crucial for applications like autonomous driving but struggle in poor visibility.
• Atmospheric turbidity significantly degrades the performance of existing image processing algorithms.
• Existing image visibility restoration methods offer solutions but can be improved.

Purpose of the Study:

• To develop a novel method for recovering clear images from those degraded by atmospheric turbidity.
• To implement a hardware accelerator for real-time image dehazing.
• To address the issue of false enlargement of white objects post-dehazing.

Main Methods:

• A supervised machine learning technique is used to estimate pixel-wise extinction coefficients.
• A novel compensation scheme is introduced to correct white object enlargement after dehazing.
• A hardware accelerator is designed and implemented on a Field Programmable Gate Array (FPGA) for real-time processing.

Main Results:

• The proposed method demonstrates superior performance compared to existing approaches on synthetic and real datasets.
• The hardware accelerator achieves a processing rate of approximately 271.67 Mpixel/s.
• Real-time processing of 4K videos at 30.7 frames per second is enabled by the accelerator.

Conclusions:

• The novel supervised learning-based method effectively restores image clarity in turbid conditions.
• The developed FPGA accelerator facilitates real-time image dehazing, crucial for practical systems.
• The approach offers a significant advancement in robust machine vision for adverse weather scenarios.