Residual Dense Network for Image Restoration.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Deep convolutional neural networks (CNNs) show success in image restoration (IR).
• Existing CNN models often underutilize hierarchical features from low-quality images, limiting performance.
• A need exists for more effective feature exploitation in deep learning-based IR.

Purpose of the Study:

• To propose a novel and efficient Residual Dense Network (RDN) for image restoration.
• To improve the exploitation of hierarchical features from all convolutional layers in IR models.
• To achieve a better balance between efficiency and effectiveness in deep IR.

Main Methods:

• Introduced Residual Dense Blocks (RDBs) for abundant local feature extraction via dense connections.
• Implemented a contiguous memory mechanism allowing direct connections between RDBs.
• Employed local and global feature fusion within RDBs to adaptively learn effective features and stabilize training.
• Applied the RDN to single image super-resolution, Gaussian image denoising, artifact reduction, and deblurring.

Main Results:

• The RDN effectively exploits hierarchical features for superior image restoration performance.
• Experiments demonstrated favorable quantitative and visual results across multiple IR tasks.
• The proposed RDN outperforms state-of-the-art methods on benchmark and real-world datasets.

Conclusions:

• The RDN offers an effective and efficient approach for image restoration by leveraging hierarchical features.
• The novel RDB architecture and feature fusion strategies contribute to improved IR performance.
• The RDN shows broad applicability and strong performance in diverse image restoration challenges.