TRG-Net: An Interpretable and Controllable Rain Generator.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Rain generation is crucial for training image processing models, but conventional methods rely on subjective physical rendering.
• Existing deep learning (DL) methods use 'closed box' generators, increasing learning difficulty and data needs.

Purpose of the Study:

• To propose a novel DL-based rain generator that incorporates the physical mechanism of rain generation.
• To explicitly encode fundamental rain factors (shape, orientation, length, width, sparsity) into a deep network.

Main Methods:

• Developed a DL-based rain generator utilizing filter parameterization for fine control over rain factors.
• The model learns rain factor distributions directly from data without requiring explicit labels.
• Employed unpaired and paired data augmentation experiments for validation.

Main Results:

• Generated rain exhibits higher quality and superior effectiveness for deraining compared to state-of-the-art methods.
• Demonstrated improved performance in both in-distribution and out-of-distribution (OOD) generalization tasks.
• Validated the model's ability to generate diverse synthetic rain samples.

Conclusions:

• The proposed rain generator effectively balances physical realism with data-driven adaptation.
• Offers enhanced controllability and adaptability for synthetic rain generation in image processing.
• Shows significant potential for improving deraining models and enabling robust OOD generalization.