Lightweight image de-snowing: A better trade-off between network capacity and performance

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Snowy conditions degrade image quality, impacting vision-based intelligent systems.
• Current de-snowing methods often have high memory consumption, limiting real-world applications.
• There is a need for efficient de-snowing solutions that balance performance and resource usage.

Purpose of the Study:

• To address the limitations of existing de-snowing methods by proposing novel networks with reduced memory footprint.
• To develop efficient and effective image de-snowing solutions for diverse application scenarios.
• To achieve a trade-off between network capacity, performance, and memory consumption.

Main Methods:

• Proposed two novel lightweight recursive networks: XLRNet and CLDRNet.
• XLRNet utilizes a single recursive strategy and two lightweight modules for resource-constrained devices.
• CLDRNet employs a dual recursive strategy and a dual coupled LSTM module (DC-LSTM) for enhanced performance.

Main Results:

• Both proposed models demonstrated effectiveness in de-snowing tasks.
• XLRNet provides a lightweight solution suitable for devices with limited memory and fast inference requirements.
• CLDRNet achieves superior de-snowing performance for devices with larger memory capacity.

Conclusions:

• The developed networks offer effective solutions for the single image de-snowing problem.
• The proposed models successfully balance performance with memory efficiency.
• These networks are suitable for various application scenarios in computer vision.