中文版本

Opportunity  

In the field of image processing, noise contamination is a pervasive issue that significantly degrades image quality, affecting applications such as medical diagnostics, aerial exploration, and underwater imaging. Traditional denoising methods rely heavily on manually tuned parameters or paired clean-noisy datasets, which are often unavailable in real-world scenarios. For instance, cameras capturing images in low-light or unstable environments (e.g., disaster zones or deep-sea explorations) produce images with unpredictable noise patterns. Existing neural network-based denoisers require ground-truth clean images for training, limiting their applicability to "blind" noise (unknown noise distributions). This patent addresses the challenge of denoising without paired training data by leveraging non-local similarity and edge-enhanced feature fusion.  

Technology  

The patent introduces a novel denoising model comprising two interconnected neural networks: a first network layer for extracting enhanced local features and a second network layer for capturing edge features via an edge detection module (e.g., Canny operator). The key innovation lies in the synergistic interaction between these layers:  

1. Non-local similarity: The model constructs a reference clean image from noisy inputs by exploiting pixel correlations, eliminating the need for paired training data.  
2. Edge-guided attention: Edge features from the second network are multiplied (point-wise) with local features from the first network, directing attention to critical structural details while suppressing noise.  
3. Heterogeneous convolutions: The first network uses mixed kernel sizes (e.g., 3×3, 3×1, 1×1) to capture multi-scale local information, while Swish/ReLU activation functions enhance nonlinear noise separation.  
4. Loss functions: Dual loss terms—one for pixel-level accuracy and another for edge preservation—are combined to optimize model robustness.  

Advantages  

• Blind noise handling: Effective for unpaired or real-world noisy images without ground-truth references.  
• Edge preservation: Explicit edge detection and fusion prevent detail loss in reconstructed images.  
• Adaptability: Heterogeneous convolutions and residual connections improve feature extraction across diverse noise types.  
• Computational efficiency: Modular design allows parallel processing of local and edge features.  

Applications  

• Medical imaging: Enhancing MRI/CT scans corrupted by sensor noise.  
• Remote sensing: Denoising satellite/aerial images affected by atmospheric interference.  
• Autonomous vehicles: Improving low-light camera feeds for navigation safety.  
• Underwater exploration: Recovering details in murky or turbulent aquatic environments.