Image prediction algorithm for foggy road scenes based on improved transformer.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Driving Systems

Background:

• Severe foggy weather significantly reduces driving visibility, posing a major risk to road safety.
• Current autonomous driving systems struggle with perception and prediction in adverse weather conditions like heavy fog.

Purpose of the Study:

• To propose a foggy image prediction algorithm for road scenes using Transformer architecture.
• To enhance the visual perception and prediction capabilities of autonomous driving systems in severe foggy weather.

Main Methods:

• Utilized a Transformer model with Taylor-expanded multi-head self-attention to reduce computational costs.
• Implemented a multi-branch architecture featuring multi-scale patch embedding with deformable convolutions for feature extraction.
• Developed an algorithm for foggy image prediction in road scenes.

Main Results:

• Achieved a Peak Signal-to-Noise Ratio (PSNR) of 12.9836 and a Structural Similarity Index Measure (SSIM) of 0.6278 on custom haze datasets.
• Demonstrated effective prediction of real images under hazy weather conditions.
• Showcased good image prediction results with relatively low computational performance requirements.

Conclusions:

• The proposed Transformer-based method effectively improves visibility in hazy conditions, addressing critical driving safety issues.
• The algorithm contributes to the advancement of autonomous driving technology by enhancing performance in adverse weather.
• The approach offers a viable solution for real-time image prediction and perception in foggy environments.