Image De-Raining Transformer.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Convolutional Neural Networks (CNNs) dominate deep learning for image de-raining.
• CNNs have limitations in capturing long-range dependencies and complex rainy artifacts due to local receptive fields.

Purpose of the Study:

• To develop an effective and efficient transformer-based architecture for image de-raining.
• To overcome the limitations of convolutional approaches in handling complex rainy patterns.

Main Methods:

• Proposed a transformer-based architecture incorporating vision task priors (locality, hierarchy) for efficient de-raining without pre-training.
• Designed complementary window-based and spatial transformers to capture both local and non-local features.
• Introduced a relative position enhanced multi-head self-attention mechanism to address positional blindness.

Main Results:

• The model demonstrates strong performance in de-raining without requiring costly pre-training.
• Achieved superior ability in capturing dependencies from both image content and spatial positions.
• Outperformed state-of-the-art methods quantitatively and qualitatively in image de-raining experiments.

Conclusions:

• The proposed transformer architecture effectively removes complex rainy artifacts.
• The model achieves superior image content recovery compared to existing de-raining techniques.
• This approach offers a promising direction for advanced image de-raining research.