Training Method for a System for De-Noising Images

Simple SummaryContent extracted from patent full text and abstract with AI.

This patent describes a new training method for an image de-noising system that uses a series of trainable bilateral filters. The system is trained with noisy images and associated 'noise maps' that indicate local noise levels in every pixel, without needing any clean, ground-truth images. By leveraging Stein's unbiased risk estimator (SURE) as a loss function, the method enables unsupervised training, facilitating effective image noise reduction in a computationally efficient way, particularly suited for medical images such as MRI or CT scans.

Use CasesContent extracted from patent full text and abstract with AI.

• Medical imaging (e.g., MRI, CT scans) to reduce noise and improve image clarity for clinical diagnoses.
• Industrial imaging for detecting defects or quality assurance where noisy images are common.
• Consumer photography or mobile devices to enhance photos taken in poor lighting conditions (low signal-to-noise ratio).
• Scientific imaging (astronomy, microscopy) where noise typically reduces the quality of acquired data.
• Security and surveillance footage enhancement.
• Image processing software and toolkits integrated into hardware systems or cloud services.

BenefitsContent extracted from patent full text and abstract with AI.

• Does not require noise-free, ground-truth images for training, enabling broader applicability and lower data acquisition barriers.
• Reduces computational cost and data requirements compared to traditional deep convolutional neural networks.
• Utilizes local noise characteristics for more precise and adaptive noise reduction per pixel.
• Preserves important image details, such as edges, thanks to bilateral filtering.
• Can be applied to various types of images and imaging modalities beyond MRI (e.g., CT, photographic images).
• Easily integrated into modern image reconstruction pipelines and hardware systems, including direct incorporation in MRI devices.
• Supports scalable and flexible training as more data becomes available, improving robustness across diverse imaging conditions.

Technical Classifications (CPCs)

Inventors & Applicants

Patent Abstract

The invention describes a training-method for training a system (12) for de-noising images, the system (12) comprising an input-interface (21), a number of trainable bilateral filters (20) designed and arranged for filtering an image provided by the input interface (21), the method comprising the steps:- providing a plurality of training images (T) as input for the system (12),- providing a number of noise maps (M) indicating the standard deviation of the noise for every pixel of a training image (T),- training the number of bilateral filters (20) based on the training images (T), the number of noise maps (M) and based on calculating analytical gradients of a loss function (L) with respect to filter parameters of the system (12), wherein at least one of the loss functions (L) is based on Stein's unbiased risk estimator.The invention further describes a related system, a filtering-method and a magnetic resonance imaging system.