Multiscale Cross-Connected Dehazing Network With Scene Depth Fusion.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Image dehazing is crucial for improving visual quality in adverse weather conditions.
• Existing methods often struggle with preserving details and achieving real-time performance.
• The correlation between hazy images and scene depth information is underexplored for dehazing.

Purpose of the Study:

• To develop an advanced image dehazing model leveraging scene depth information.
• To propose a multiscale network architecture for enhanced feature extraction.
• To improve the accuracy and visual quality of dehazed images.

Main Methods:

• A multiscale cross-connected dehazing network integrating scene depth.
• Separate encoding and decoding paths for hazy and depth images with cross-connections.
• Utilized input pyramid for multi-level feature extraction and wavelet pooling and residual channel attention modules (RCAMs).

Main Results:

• The proposed model effectively fuses hazy and depth image information.
• Ablation studies confirmed the efficacy of wavelet pooling and RCAMs.
• Achieved superior performance over state-of-the-art methods in PSNR, SSIM, and visual quality.

Conclusions:

• The multiscale cross-connected network with depth fusion offers a robust solution for image dehazing.
• The integration of depth information significantly boosts dehazing performance.
• The model demonstrates potential for real-world applications requiring high-quality image restoration.