A raw data modeling noise combined with real pairing data training method
By training with a mixture of real data from different individual cameras of the same model and noise data from physical modeling, the generalization problem of neural network denoising methods in extremely dark scenes was solved, improving image quality and model applicability.
CN122265068APending Publication Date: 2026-06-23HEFEI JUNZHENG TECH CO LTD
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HEFEI JUNZHENG TECH CO LTD
- Filing Date
- 2024-12-20
- Publication Date
- 2026-06-23
AI Technical Summary
Technical Problem
Existing neural network denoising methods require a large amount of paired data in extremely dark scenes, and the noise patterns and intensities of different individual cameras of the same model vary, resulting in limited denoising performance and an inability to generalize to all cameras.
Method used
Real data from different individual cameras of the same model is collected and combined with noise data from physical modeling. The model is trained using a mixed training set, and the real data is used to improve image details and colors to construct the training set.
Benefits of technology
It improves denoising performance on different cameras, enhances image quality in extremely dark scenes, and strengthens the model's generalization ability.
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN122265068A_ABST
Abstract
The application provides a raw data modeling noise combined with a real paired data training method, which is an improvement of the collected paired data, and details and colors of an image are improved by adding a part of real paired data; the method comprises the following steps: S1, a real data set collection process; and S2, a process of synthesizing a false data set. In order to obtain a model with good denoising effect on different individual cameras and suitable for various camera devices, first, same models and different individual cameras are prepared, then a certain amount of real data is collected by using the cameras under different light conditions and scenes, the real data is processed into a paired data training set, then the data is mixed for training by combining physical modeling, and finally the model is generalized to other cameras. The paired data of the same model and different cameras are used to supplement the noise of the part not correctly modeled by the physical modeling, the noise problem of normal light and very short exposure time is improved, and the model can obtain good denoising effect in the same model and different cameras.
Need to check novelty before this filing date? Find Prior Art