Vision Transformers for Single Image Dehazing.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Image dehazing is crucial for low-level vision, traditionally dominated by convolutional neural networks.
• Vision Transformers show promise in high-level vision but haven't been effectively adapted for image dehazing.

Purpose of the Study:

• To adapt Vision Transformers for image dehazing by addressing limitations of existing architectures like Swin Transformer.
• To propose an improved Transformer model, DehazeFormer, for enhanced image dehazing.

Main Methods:

• Modified Swin Transformer architecture with tailored normalization layers, activation functions, and spatial information aggregation.
• Training and evaluation of multiple DehazeFormer variants on diverse datasets, including a newly collected remote sensing dataset.

Main Results:

• The small DehazeFormer model surpasses FFA-Net in performance with significantly lower parameters and computational cost.
• The large DehazeFormer model achieves the first-ever PSNR over 40 dB on the SOTS indoor dataset, setting a new state-of-the-art.

Conclusions:

• DehazeFormer offers a superior Transformer-based solution for image dehazing, outperforming existing methods in efficiency and effectiveness.
• The proposed model demonstrates strong generalization capabilities, particularly for challenging real-world scenarios like remote sensing image dehazing.