Synthetic Training Data Generation for Low-Light Image Enhancement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for generating training data for image enhancement models are time-consuming and impractical, especially for capturing images in low light conditions or with motion, as they require physically capturing input and output images, which can result in blurred or unfeasible data sets.
Innovation Solution
The system generates input images based on target images using predetermined pixel values, allowing for the simulation of images captured in low light conditions, thereby saving time and enabling the use of scenes that would otherwise be difficult to capture, such as those with motion, by creating a mapping between illumination data of reference and input images and applying supervised learning algorithms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If physically capturing input and output images is used to generate training data, then the training data reflects real capture conditions, but the process is time-consuming and impractical for low light conditions or motion scenes
Solution Approach 1:
The patent creates synthetic training data by copying and transforming existing output images into corresponding input images. Instead of capturing both input and output images physically, the system generates input images by applying simulated degradation models (darkness, noise, blur) to output images, thereby replicating the training data without time-consuming physical capture processes
Solution Approach 2:
The patent performs preliminary actions by pre-computing degradation models and characteristics before actual training. The system pre-processes output images to create corresponding input images with simulated low-light conditions, so that when training occurs, the data is already prepared and no time-consuming capture processes are needed
2Measurement precision
If physically capturing input images with longer exposure times is used, then better low light images can be obtained, but motion scenes become blurred and unusable
Solution Approach 1:
The patent performs preliminary actions by pre-computing degradation models and characteristics before actual training. The system pre-processes output images to create corresponding input images with simulated low-light conditions, so that when training occurs, the data is already prepared and no time-consuming capture processes are needed
Solution Approach 2:
The patent creates synthetic training data by copying and transforming existing output images into corresponding input images. Instead of capturing both input and output images physically, the system generates input images by applying simulated degradation models (darkness, noise, blur) to output images, thereby replicating the training data without time-consuming physical capture processes
3Ease of manufacture
If conventional image capture hardware is used for training data generation, then existing devices can be utilized, but specialized hardware is needed to capture low light and motion scenes effectively
Solution Approach 1:
The patent creates synthetic training data by copying and transforming existing output images into corresponding input images. Instead of capturing both input and output images physically, the system generates input images by applying simulated degradation models (darkness, noise, blur) to output images, thereby replicating the training data without time-consuming physical capture processes
Solution Approach 2:
The patent replaces the mechanical capture system with a computational system. Instead of using physical cameras and optical systems to capture low-light and motion scenes, the system uses software-based degradation models that mathematically simulate these conditions, substituting mechanical capture with computational generation
Data Source
AI summary
Systems and methods for generating training data using synthesized input images for training a machine learning model for image enhancement in accordance with embodiments of the invention are described. The system may access a target image (e.g., captured by an imaging device), and generate an input image that corresponds to the target image. The input image and target image may then be used (e.g., as part of a training data set) to train the machine learning model. For example, a generated input image corresponding to a target image may represent a version of the target image as if it were captured in low light. The target image may be a target illuminated output to be generated by enhancing the input image. The input image and target image may be used to train a machine learning model to enhance images such as those captured in low light.