Image dehazing algorithm based on deep transfer learning and local mean adaptation.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Haze significantly degrades image quality and reduces visual perception range, impacting various applications.
• Existing image dehazing methods struggle with bright regions and exhibit weak noise resistance, leading to artifacts and low peak signal-to-noise ratio (PSNR) values.
• Simultaneously achieving effective dehazing in bright areas and robust noise suppression remains a challenge.

Purpose of the Study:

• To propose a novel image dehazing algorithm addressing limitations in bright regions and noise resistance.
• To improve dehazing performance, enhance image details, and ensure high-quality visual output.
• To develop a robust and generalizable dehazing solution applicable to diverse datasets and real-world scenarios.

Main Methods:

• A deep transfer learning-based atmospheric light estimation module.
• A local mean adaptation technique for transmission map estimation.
• Integration of haze-free image reconstruction, image enhancement, and noise reduction modules.

Main Results:

• The proposed algorithm achieves superior dehazing performance across four diverse datasets (Self-Made Synthetic Hazy, SOTS, NH-HAZE, O-HAZE).
• Dehazed images exhibit no color distortion, with PSNR values consistently exceeding 30 dB and Structural Similarity Index Measure (SSIM) over 85%.
• The method effectively handles bright regions and significantly reduces residual noise, demonstrating strong generalization capabilities.

Conclusions:

• The proposed deep transfer learning and local mean adaptation-based dehazing algorithm offers a significant advancement over existing methods.
• The model provides high-quality, artifact-free dehazed images with enhanced noise resistance and strong generalization.
• The developed framework has potential applications in autonomous driving, intelligent surveillance, and other vision-based systems.