Recurrent Neural Network Image Denoising Without Ground Truth
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
High-resolution imaging techniques like FIB-SEM produce images with omnipresent noise, making it difficult to obtain noise-free images due to uncontrollable optical and thermal effects, and existing machine learning models struggle to denoise images when ground truth images are unavailable.
Innovation Solution
A machine learning model with a modular architecture and a recurrent neural network (RNN) is trained using sequences of noisy images to identify and remove noise iteratively, utilizing building units and noise attention blocks to denoise images without requiring ground truth pairs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-resolution imaging techniques are used, then image resolution is improved, but noise increases
Solution Approach 1:
The patent extracts and removes noise from high-resolution images using machine learning models. The model identifies and separates noise components from the actual image data, effectively removing the harmful noise while preserving the high-resolution details of the original image.
Solution Approach 2:
The patent introduces machine learning models as an intermediary between the noisy high-resolution image and the final cleaned image. This intermediary system processes the image through multiple layers of neural networks to transform the noisy input into a clean output without directly modifying the imaging technique itself.
2Measurement precision
If machine learning models are trained with ground truth pairs, then denoising accuracy is improved, but data requirements and complexity increase
Solution Approach 1:
The patent implements self-service training where the machine learning model generates its own training data. The model creates synthetic noisy-c clean image pairs by adding controlled noise to ground truth images, eliminating the need for extensive manual annotation and ground truth pair collection while maintaining high denoising accuracy.
Solution Approach 2:
The patent performs preliminary generation of training data before actual model training. By pre-generating synthetic training pairs with known ground truth, the system prepares adequate training material in advance, simplifying the subsequent training process and reducing the complexity of data collection and annotation.
Data Source
AI summary
In some examples, a machine learning model may be trained to denoise an image. In some examples, the machine learning model may identify noise in an image of a sequence based at least in part, on at least one other image of the sequence. In some examples, the machine learning model may include a recurrent neural network. In some examples, the machine learning model may have a modular architecture including one or more building units. In some examples, the machine learning model may have a multi-branch architecture. In some examples, the noise may be identified and removed from the image by an iterative process.