Deep guided transformer dehazing network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Deep learning models have advanced single image dehazing.
• Convolutional neural networks (CNNs) in dehazing are limited by local receptive fields.
• Capturing long-range dependencies is crucial for accurate haze density estimation.

Purpose of the Study:

• To develop a novel deep learning model for single image dehazing.
• To address the limitations of local feature extraction in existing CNN-based methods.
• To enhance both the accuracy and speed of image dehazing.

Main Methods:

• A novel Deep Guided Transformer Dehazing Network (DGTRAN) was designed.
• A transformer-based subnetwork was employed to capture long-range dependencies for global haze information.
• A CNN subnetwork was utilized for local detail restoration, and a guided filter was integrated to accelerate processing.

Main Results:

• The proposed DGTRAN model demonstrated superior performance on both natural and simulated hazy images.
• Experimental results showed consistent improvements over state-of-the-art dehazing methods.
• The integration of the guided filter effectively improved the dehazing speed without compromising quality.

Conclusions:

• The DGTRAN model successfully overcomes the local limitations of traditional convolutional approaches in image dehazing.
• Combining transformers for global context and CNNs for local details provides a robust solution for effective and efficient dehazing.
• The proposed method offers a significant advancement in single image dehazing research.