Multi-level perception fusion dehazing network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Image dehazing is crucial for AI systems, but current models suffer from limited receptive fields and information loss during feature extraction.
• Existing dehazing methods often produce incomplete or structurally flawed results due to these limitations.

Purpose of the Study:

• To propose a novel multi-level perception fusion dehazing network (MPFDN) to address the shortcomings of existing image dehazing techniques.
• To improve feature extraction, expand perceptual fields, and preserve spatial background information for enhanced dehazing outcomes.

Main Methods:

• Developed a multi-level perception fusion dehazing network (MPFDN) integrating features across scales.
• Implemented an error feedback mechanism and feature compensator to mitigate feature loss.
• Obtained high-quality dehazed images by subtracting the generated residual image from the original hazy image.

Main Results:

• The MPFDN effectively integrates multi-scale features, expands the network's perceptual field, and extracts comprehensive spatial background information.
• The error feedback and feature compensation mechanisms successfully addressed feature loss during the dehazing process.
• Extensive experiments confirmed the outstanding performance of the proposed method on both synthetic and non-homogeneous haze datasets.

Conclusions:

• The proposed MPFDN significantly improves image dehazing by addressing limitations in feature extraction and information preservation.
• The method demonstrates robust performance across diverse datasets, offering a superior solution for AI-driven image recognition and classification.