Self-Residual Blind Denoising for Texture-Preserving Noise Removal
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional self-supervised blind denoising methods fail to effectively eliminate spatially correlated noise in real-world images without degrading texture details, as they rely on Pixel-shuffle Downsampling (PD) which destroys high-frequency details like edges.
Innovation Solution
A blind denoising device and method that uses a self-residual learning approach, employing a receiving unit, PD unit, predicted image producing unit, and learning unit to eliminate noise without PD, utilizing self-supervised losses to enhance texture preservation and denoising performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If Pixel-shuffle Downsampling (PD) is used to eliminate spatial correlation of noise, then noise elimination performance is improved, but texture details are degraded
Solution Approach 1:
The model segments the denoising task into two distinct networks: a first network that processes the original noisy image to preserve texture details, and a second network that processes downsampled images to eliminate spatially correlated noise. By dividing the single denoising function into separate specialized networks, the system achieves both noise elimination and texture preservation without requiring PD during inference.
Solution Approach 2:
The invention operates in the frequency domain by processing images at different resolutions. The second network works on downsampled (lower resolution) images to capture global noise patterns, while the first network works on the original (higher resolution) image to preserve local texture details. This multi-scale approach allows the model to address spatial correlation without degrading texture.
2Manufacturing precision
If a denoising model is trained without PD process to preserve texture details, then texture preservation is improved, but noise elimination capability may be insufficient
Solution Approach 1:
The invention merges the outputs of two specialized networks: the first network that preserves texture details from the original image, and the second network that eliminates spatially correlated noise from downsampled images. By combining these complementary results, the system achieves comprehensive noise elimination while maintaining texture fidelity, solving the limitation of training without PD.
Solution Approach 2:
The downsampled images serve as an intermediary representation that captures spatial correlation patterns without containing fine texture details. The second network processes this intermediary representation to learn noise patterns, which are then applied to enhance the output of the first network, achieving effective noise elimination without directly degrading the original image's texture.
3Reliability
If multiple networks are used for noise elimination from original and downsampled images, then denoising performance is improved, but device complexity increases
Solution Approach 1:
Both the first and second networks use the same blind-spot network (BSN) architecture, making the network design universal and reusable. This multi-functionality allows the same architectural blueprint to serve different purposes: the first network handles texture preservation while the second handles noise pattern learning, reducing overall system complexity through architectural consistency.
Solution Approach 2:
The second network is essentially a copied version of the first network's architecture (both use BSN), but applied to downsampled images. This copying approach allows the system to leverage the same proven architecture for a complementary task without designing a completely new complex structure, thereby improving performance while controlling complexity through architectural reuse.
Data Source
AI summary
The present invention relates to a blind denoising device including: a receiving unit for receiving an original noisy image; a Pixel-shuffle Downsampling (PD) unit for performing PD for the original noisy image to produce at least one or more downsampled images; a predicted image producing unit for eliminating the noise from the original noisy image and the downsampled images to produce at least one or more predicted images restored to the shape of the original noisy image; and a learning unit for performing optimized learning for the predicted image producing unit, based on at least one or more self-supervised losses of the predicted images and the original noisy image.