Semi-Supervised Image Deraining Using Gaussian Processes.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Convolutional Neural Network (CNN)-based image deraining methods excel but require fully labeled data.
• Real-world deraining datasets are challenging to acquire, leading to models trained on synthetic data that generalize poorly.
• Utilizing real-world data for training image deraining networks remains underexplored.

Purpose of the Study:

• To develop a semi-supervised learning framework for image deraining that effectively utilizes unlabeled real-world data.
• To improve the generalization capability of image deraining networks trained on synthetic datasets.
• To enhance the performance of deraining models by incorporating unlabeled real-world images.

Main Methods:

• A Gaussian Process-based semi-supervised learning framework is proposed.
• Latent space vectors of unlabeled data are modeled using Gaussian Processes.
• Pseudo-ground-truth is generated from Gaussian Processes to supervise the network on unlabeled data at intermediate levels.

Main Results:

• The proposed method effectively leverages unlabeled data, leading to significantly better performance than labeled-only training.
• Experiments on challenging datasets (Rain800, Rain200L, DDN-SIRR) demonstrate superior performance.
• Incorporating unlabeled real-world images in the GP-based framework outperforms existing methods.

Conclusions:

• The Gaussian Process-based semi-supervised framework enables effective image deraining by bridging the gap between synthetic and real-world data.
• The approach significantly enhances deraining performance and generalization capabilities.
• This method offers a promising direction for training robust image deraining models using readily available unlabeled real-world data.