Leveraging Frequency Analysis for Image Denoising Network Pruning.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Network pruning is a common technique for model compression, reducing storage and computational costs.
• Existing pruning methods are primarily designed for high-level vision tasks and are unsuitable for low-level tasks like image denoising.
• Norm-based pruning criteria fail for image denoising due to differing feature granularities and network objectives.

Purpose of the Study:

• To develop a novel pruning method specifically for image denoising networks.
• To address the limitations of existing pruning techniques in low-level vision tasks.
• To enhance the performance and interpretability of pruned image denoising models.

Main Methods:

• Proposed a new filter evaluation method: High-Frequency Components Pruning (HFCP).
• HFCP assesses filter importance by analyzing high-frequency components within the network.
• Validated HFCP across four mainstream image denoising networks.

Main Results:

• HFCP is the first pruning method specifically designed for image denoising.
• The method is straightforward, applicable to various noise types, and enhances high-frequency information content.
• HFCP improves the pruned model's ability to distinguish signal from noise reliably and interpretably.

Conclusions:

• High-Frequency Components Pruning (HFCP) is an effective and specialized technique for image denoising network compression.
• HFCP overcomes the limitations of traditional pruning methods in low-level vision tasks.
• This novel approach offers improved performance and interpretability for denoising models.