Neural Network Training for Low-Light Image Denoising
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Low-light images captured by cameras often exhibit poor contrast and significant noise due to low photon counts, and existing image processing techniques struggle to accurately reproduce the scene, especially in extreme low-light conditions, as they require careful parameter selection and may introduce artifacts or noise.
Innovation Solution
A method for training neural networks by selecting representative images with high signal-to-noise ratio (SNR) and lower SNR input images, where the number of saturated pixels is used to filter out unsuitable images, allowing the network to learn rules for denoising low-light images across a wide range of capture conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If algorithmic filters are used for image processing, then image quality can be improved within a narrow range of capture conditions, but the effectiveness is limited when capture conditions fall outside the trained parameter range
Solution Approach 1:
The patent changes the fundamental parameter of image processing from fixed algorithmic filters to neural networks trained on diverse capture conditions. The neural network learns to handle varying exposure, noise levels, and lighting conditions through training on multiple images with different parameters, enabling adaptive processing across wide ranges of capture conditions.
Solution Approach 2:
The neural network model serves multiple functions by being trained on a diverse dataset encompassing various capture conditions (different exposures, noise levels, lighting). This universal training enables the single model to effectively process images across a broad spectrum of conditions, replacing multiple condition-specific algorithms.
2Illumination intensity
If increasing exposure time is used to improve low-light image quality, then brightness is improved, but motion blur artifacts are introduced
Solution Approach 1:
The patent converts the harmful effect of motion blur and noise in low-light images into beneficial training data. By training neural networks on images with varying degrees of noise and blur, the network learns to recognize and reconstruct the underlying scene, effectively using the harmful artifacts as learning opportunities to improve denoising and deblurring capabilities.
Solution Approach 2:
The patent replaces the mechanical approach of increasing exposure time to improve brightness with a computational approach using neural networks. Instead of physically extending exposure to capture more light, the system uses learned models to reconstruct high-quality images from low-light, noisy inputs through intelligent algorithms.
3Illumination intensity
If increasing photosensor gain is used to improve low-light image quality, then brightness is improved, but noise is amplified
Solution Approach 1:
The patent converts the harmful noise amplification effect into a beneficial training opportunity. By training neural networks on images with known noise characteristics and ground truth labels, the network learns to distinguish between signal and noise, effectively using the amplified noise as training data to improve denoising capabilities.
Solution Approach 2:
The patent replaces the mechanical increase of photosensor gain with a computational neural network approach. Instead of physically amplifying the signal and accepting the accompanying noise, the system uses learned models to reconstruct the original signal from the noisy input, effectively substituting physical amplification with intelligent signal recovery.
4Measurement precision
If manual selection of ground truth and input image pairs is performed, then training accuracy can be controlled, but computational intensity and time consumption increase
Solution Approach 1:
The patent implements self-service through automated image selection algorithms that independently identify suitable ground truth and input image pairs without human intervention. The system automatically evaluates image quality metrics, selects appropriate pairs, and prepares training data, freeing computational resources and time from manual curation while maintaining training accuracy.
Solution Approach 2:
The patent performs preliminary automated selection and preprocessing of training images before the actual training begins. By pre-filtering and organizing images into appropriate pairs based on quality metrics and scene similarity, the system reduces the computational burden during training and accelerates the overall process while maintaining high accuracy.
Data Source
AI summary
This disclosure provides methods, devices, and systems for low-light imaging. The present implementations more specifically relate to selecting images that can be used for training a neural network to infer denoised representations of images captured in low light conditions. In some aspects, a machine learning system may obtain a series of images of a given scene, where each of the images is associated with a different SNR (representing a unique combination of exposure and gain settings). The machine learning system may identify a number of saturated pixels in each image and classify each of the images as a saturated image or a non-saturated image based on the number of saturated pixels. The machine learning system may then select the non-saturated image with the highest SNR as the ground truth image, and the non-saturated images with lower SNRs as the input images, to be used for training the neural network.