Neural Network Underexposure Correction for Low Light Noise Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Low light environments pose challenges for camera image capture due to noise, requiring either increased sensor sensitivity, larger apertures, longer exposure times, or artificial lighting, each with associated drawbacks such as noise amplification, costly equipment, motion blur, and unnatural overexposure.
Innovation Solution
A system utilizing a convolutional neural network to process underexposed images, reducing noise and improving image quality with minimal loss, by determining ideal exposure settings, capturing underexposed images, and then processing them to adjust camera operations such as ISO and shutter speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Illumination intensity
If sensor sensitivity is increased to improve image brightness in low light, then image brightness is improved, but sensor noise is amplified
Solution Approach 1:
The patent replaces traditional optical/mechanical noise reduction methods (aperture adjustments, exposure time extensions) with a neural network-based digital processing system. The neural network learns to distinguish between signal and noise patterns, substituting physical camera adjustments with intelligent algorithmic processing to reduce noise while preserving image brightness.
Solution Approach 2:
The patent changes the processing parameters from traditional camera settings (aperture, shutter speed, ISO) to neural network training parameters and digital processing thresholds. By training the neural network on various exposure conditions and noise patterns, the system optimizes digital parameters to reduce noise while maintaining image quality, rather than relying on physical parameter adjustments that have diminishing returns.
2Illumination intensity
If aperture size is increased to improve light gathering, then image brightness is improved, but lens size, weight, and cost increase
Solution Approach 1:
The patent substitutes the mechanical aperture adjustment system with a neural network-based image processing system. Instead of physically opening the aperture to gather more light, the system captures underexposed images and uses neural networks to reconstruct the image quality, eliminating the need for large, heavy lenses while achieving similar or better image quality.
Solution Approach 2:
The patent creates a computational copy of the ideal image from the underexposed original. The neural network learns to synthesize a high-quality image representation from the low-light input, effectively copying the desired image characteristics without requiring the physical optical system to perform the impossible task of gathering enough light with a small aperture.
3Illumination intensity
If exposure time is extended to improve light collection, then image brightness is improved, but motion blur and camera shake artifacts increase
Solution Approach 1:
The patent performs preliminary action by capturing the image data quickly underexposed, before motion blur can occur. The neural network then processes this rapid capture to reconstruct the image, effectively reversing the traditional approach of waiting for sufficient light accumulation and instead using intelligent processing to enhance the quickly captured underexposed data.
Solution Approach 2:
The patent replaces the mechanical exposure time adjustment with neural network-based image reconstruction. Instead of extending the shutter open time to gather light (which causes motion blur), the system uses a neural network to computationally reconstruct the image from brief, underexposed captures, substituting physical time extension with digital processing.
4Illumination intensity
If flash lighting is used to improve low light image quality, then image brightness is improved, but cost, deployment difficulty, and unnatural appearance increase
Solution Approach 1:
The patent substitutes the physical flash lighting system with a neural network-based processing system. Instead of deploying complex flash hardware and controlling illumination timing, the system captures images in ambient light and uses neural networks to enhance the image quality, dramatically simplifying deployment and eliminating the need for additional lighting equipment.
Solution Approach 2:
The patent creates a computational copy of the ideal exposure result from the ambient light capture. The neural network learns to synthesize what the image would look like under proper exposure conditions, effectively copying the desired lighting characteristics without requiring actual flash illumination, thereby reducing deployment complexity and maintaining natural appearance.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method for image capture includes determining an exposure range and setting at least one camera parameter to capture an underexposed image outside the exposure range. The underexposed image is processed using a neural network to recover image details. Image defects due to camera or object motion blur can be reduced.