Beam Splitter for Simultaneous Multi-Exposure Image Capture
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning systems face issues with unwanted artefacts like motion blur and distribution shift when training on image pairs captured with different exposure settings, and perspective changes when used in applications like autonomous vehicles or stereoscopic vision systems.
Innovation Solution
A method and device that split a common light path using a beam splitter to capture simultaneous image pairs with the same perspective, allowing for automatic generation of training data without artefacts, using identical or similar optical and image sensors, and optionally incorporating modification filters to simulate real-world conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If sequential photographing with different exposure settings is used, then training data for image enhancement can be obtained, but motion blur and other artefacts are introduced in the target image
Solution Approach 1:
The patent combines multiple exposure settings within a single photographing operation using a beam splitter to divide the light path. This allows simultaneous capture of images with different exposure settings from the same scene at the same moment, eliminating motion blur while maintaining the ability to train image enhancement models.
Solution Approach 2:
The patent introduces a beam splitter as an intermediary optical element to separate the light path into multiple paths with different exposure settings. This mediator enables the system to capture multiple exposure variants simultaneously without sequential photographing, thus avoiding motion blur artifacts.
2Quantity of substance
If sequential photographing is used, then training data can be obtained, but distribution shift occurs due to objects entering or leaving the scene
Solution Approach 1:
The patent merges multiple exposure settings into a single simultaneous photographing operation. This ensures that all training images are captured from the exact same scene configuration at the same moment, eliminating distribution shift caused by moving objects while still providing diverse training data through different exposure settings.
3Loss of time
If side-by-side camera mounting is used, then simultaneous image pairs can be captured, but perspective differences are introduced
Solution Approach 1:
The patent uses a single camera with a beam splitter to combine multiple light paths instead of mounting cameras side-by-side. This ensures that all images are captured from the exact same optical center and perspective while still achieving simultaneous capture with different exposure settings, thus eliminating perspective differences.
4Manufacturing precision
If manual generation of example data is used, then data quality can be ensured, but the process is tedious and time consuming
Solution Approach 1:
The patent enables automatic generation of training data by capturing images with different exposure settings simultaneously through a beam splitter system. This self-service approach eliminates the need for manual data preparation while ensuring consistent quality, as the system automatically captures matched image pairs from the same scene.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
The solution enables the generation of high-quality machine learning training data without distribution shift or perspective changes, facilitating effective training of systems for image processing and object recognition, especially in security-critical applications like autonomous vehicles.
Implementation Method 1
a common light path is split by a beam splitter into a first light path and a second light path
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a method and a device for generating machine learning training data. According to the method, corresponding image pairs are captured, wherein each image pair comprises a first image and a second image. To do so, a common light path (3) is split by a beam splitter (4) into a first light path (5) to a first optical arrangement (7) and a second light path (6) to a second optical arrangement (8), wherein the first optical arrangement (7) and the second optical arrangement (8) differ in at least one distinguishing feature. The first image is provided by a first image sensor (9) of the first optical arrangement (7) and the second image is provided simultaneously to the first image by a second image sensor (11) of the second optical arrangement (8). Finally, the first image and the second image are associated with one another and are saved to a non-volatile memory. The invention further relates to a method for training a machine learning system.