Machine-Learning HDR Image Construction with Patterned Illumination
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional microscopy techniques face challenges in capturing high-dynamic range images due to limited dynamic range, resulting in glare, dark spots, and difficulty in differentiating subtle contrasts, which hinders accurate analysis.
Innovation Solution
Employing Fourier ptychographic microscopy and machine learning to construct high-dynamic range images without precise exposure optimization, using a training set of images illuminated with multiple patterns and a machine learning model trained with ground truth images to correlate and enhance image detail capture.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Illumination intensity
If conventional microscopy techniques are used to capture images, then the imaging process is simple and fast, but the dynamic range is limited resulting in glare and dark spots
Solution Approach 1:
The imaging process is segmented into multiple exposure captures (underexposed, properly exposed, and overexposed images) that are subsequently combined through computational processing to achieve extended dynamic range while maintaining operational simplicity
Solution Approach 2:
A computational processing step acts as an intermediary between the captured exposure images and the final HDR image, combining multiple exposures to overcome the limited dynamic range of conventional microscope cameras without requiring hardware modifications
2Illumination intensity
If multiple images with different exposure settings are combined to form HDR image, then the dynamic range is increased, but the process requires optimization of exposure range which varies between samples
Solution Approach 1:
The system performs self-service by automatically determining the optimal exposure range for each sample through a preliminary scan, eliminating the need for manual optimization and making the HDR imaging process sample-adaptive and operationally simple
Solution Approach 2:
A preliminary scan is performed before the actual HDR image acquisition to automatically determine the appropriate exposure settings for the specific sample, preparing the optimal exposure range in advance to simplify the subsequent imaging process
3Manufacturing precision
If conventional microscopy is used, then the equipment is simple, but subtle contrast cannot be rendered and colors appear less vivid
Solution Approach 1:
Computational processing acts as an intermediary that enhances image details, contrast, and color vividness by combining information from multiple exposure images, achieving high manufacturing precision without requiring complex hardware modifications
Solution Approach 2:
The solution transitions from a single-dimension (single exposure) imaging approach to a multi-dimensional approach by capturing and combining images at different exposure levels, thereby extracting enhanced detail information that cannot be obtained with conventional single-exposure microscopy
Data Source
Figure 1A
Figure 1B~2
Figure 3A~3F
AI summary
The present inventive concept relates to a method (30) for training a machine learning model to construct a high-dynamic range image depicting a sample. The method (30) comprises: acquiring (S300) a training set of images by: illuminating (S302), using an illumination system (110), the training sample (142) with a plurality of illumination patterns, and capturing (S304) an image of the training sample (142) for each illumination pattern of the plurality of illumination patterns; acquiring (S306) a high-dynamic range image depicting the training sample (142), wherein the high-dynamic range image has a dynamic range higher than a dynamic range of an image of the training set of images; forming (S308) a ground truth comprising the high-dynamic range image; and training (S310) the machine learning model to construct the high-dynamic range image depicting a sample using the training set of images and the ground truth.