Driving scene image Dehazing model based on multi-branch and multi-scale feature fusion

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Image dehazing is crucial for improving visual quality in critical applications like autonomous driving and surveillance.
• Existing methods often struggle with complex scenes and real-time processing demands.

Purpose of the Study:

• To develop an innovative and efficient image dehazing model for enhanced visual perception.
• To address limitations in current dehazing techniques, particularly for real-world driving scenarios.

Main Methods:

• A multi-branch and multi-scale feature fusion network combining local and global features using depthwise separable convolutions and state space models (Mamba).
• Integration of spatial and frequency domain information via convolutional layers and Fourier transforms.
• A feature fusion mechanism employing channel attention and residual connections for dynamic feature weighting.

Main Results:

• The proposed model achieves superior dehazing performance on real-world foggy driving datasets.
• Demonstrates high processing efficiency with a frame rate of 32.41 FPS and 67.96 MPx/s.
• Successfully combines Mamba and convolution for efficient global feature extraction in driving scenes.

Conclusions:

• The novel model offers reliable visual perception and improved adaptability in complex environments.
• The fusion of spatial, frequency, and advanced network architectures enables effective and efficient haze removal.
• This work represents a significant advancement in real-time image dehazing for autonomous systems.