CMID: Crossmodal Image Denoising via Pixel-Wise Deep Reinforcement Learning

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Image noise reduction is essential for many computer vision tasks.
• Current denoising methods lack cross-modality generalization, limiting their applicability.
• Existing techniques often focus on specific noise types, hindering broad performance.

Purpose of the Study:

• To develop a pixel-wise, crossmodal image-denoising method using deep reinforcement learning.
• To enhance the generalization performance of image denoising across different modalities.
• To mimic the iterative, step-wise image processing approach used by human experts.

Main Methods:

• A deep reinforcement learning framework was employed for pixel-wise image denoising.
• A novel similarity reward function was introduced to guide the learning process.
• An expanded action set was designed to address multiple noise types within a unified framework.
• The method was trained and evaluated on RGB, infrared, and terahertz image datasets.

Main Results:

• The proposed method demonstrated superior performance in crossmodal image denoising compared to state-of-the-art techniques.
• Experiments confirmed the effectiveness of the similarity reward in optimizing the denoising sequence.
• The designed action space successfully enabled the handling of diverse noise characteristics across modalities.
• The method achieved significant improvements on publicly available datasets.

Conclusions:

• The deep reinforcement learning approach offers a powerful solution for generalized crossmodal image denoising.
• The method effectively models human-like iterative processing for noise removal.
• This work advances the capability of AI in handling complex image noise across various sensing modalities.