Neural Network Low-Light Image Denoising via Quantile Scaling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Low-light images captured by image capture devices often exhibit poor contrast and significant noise due to low photon counts, and existing image processing techniques struggle to effectively denoise such images, especially those with short exposure times.
Innovation Solution
The method involves receiving pixel values representing a low-light image, determining a scaling factor based on the kth quantile of the pixel values' range, and performing a scaling operation to normalize the pixel values. These normalized pixel values are then input into a neural network trained to reduce noise and reproduce a denoised image.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If algorithmic filters are used for image processing, then image quality can be improved within a narrow range of capture conditions, but the technique is not suitable for extremely low light conditions with short exposure times
Solution Approach 1:
The patent transforms the image processing approach by changing from fixed algorithmic parameters to adaptive neural network parameters. The neural network is trained on diverse low-light images with varying exposure times, illuminance levels, and noise characteristics, enabling it to adapt to different capture conditions rather than being limited to a narrow parameter range.
Solution Approach 2:
The patent creates a learned mapping model (neural network) that copies the relationship between noisy low-light images and their denoised versions from training data. This learned model can then be applied to new images without requiring manual parameter adjustment for each capture condition.
2Reliability
If traditional denoising algorithms are applied to low-light images, then noise reduction can be achieved, but computational intensity is high
Solution Approach 1:
The patent performs denoising operations in advance during the training phase, where the neural network learns optimal denoising transformations from paired training images. Once trained, the network can rapidly apply learned denoising patterns to new images without performing intensive computational operations during actual image processing.
Solution Approach 2:
The patent replaces traditional mechanical denoising algorithms (which perform iterative mathematical operations) with a neural network model that uses learned feature representations. This substitution transforms computationally intensive algorithmic processing into more efficient pattern recognition based on pre-learned relationships.
Data Source
AI summary
This disclosure provides methods, devices, and systems for low-light imaging. In some implementations, an image processor may be configured to reduce or remove noise associated with an image based, at least in part, on a neural network. For example, the neural network may be trained to infer a denoised representation of the image. In some aspects, the image processor may scale the brightness level of the image to fall within a normalized range of values associated with the neural network. In some other aspects, a machine learning system may scale the brightness levels of input images to match the brightness levels of ground truth images used to train the neural network. Still further, in some aspects, the machine learning system may scale the brightness levels of the input images and the brightness levels of the ground truth images to fall within the normalized range of values during training.