An Efficient Dehazing Algorithm Based on the Fusion of Transformer and Convolutional Neural Network

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Single image dehazing is crucial for restoring clarity and color in degraded images.
• Deep learning has significantly advanced image restoration tasks, including dehazing.
• Transformer models show promise in computer vision but have limitations when applied alone to dehazing.

Purpose of the Study:

• To develop a novel Transformer-Convolution fusion dehazing network (TCFDN) for enhanced single image dehazing.
• To leverage the complementary strengths of Transformer and Convolutional Neural Networks (CNNs) for improved haze removal.
• To address the limitations of existing Transformer-based and CNN-based dehazing algorithms.

Main Methods:

• Proposed a Transformer-Convolution fusion dehazing network (TCFDN) utilizing a self-encoder structure.
• Introduced a Transformer-Convolution hybrid layer with an adaptive fusion strategy.
• Integrated Swin-Transformer for global feature modeling and CNNs for local feature extraction and reconstruction.

Main Results:

• The TCFDN demonstrated superior performance compared to existing advanced dehazing algorithms.
• Experiments validated the effectiveness of combining global and local feature modeling for haze removal.
• Ablation studies confirmed the contribution of the proposed hybrid layer and fusion strategy.

Conclusions:

• The proposed TCFDN effectively enhances image dehazing capabilities by fusing global and local feature extraction.
• The adaptive fusion strategy within the hybrid layer is key to maximizing the benefits of both Transformer and CNN components.
• This work provides a robust framework and theoretical evidence for hybrid deep learning approaches in image restoration.