Electron Microscopy Reconstruction With Pre-Learned Dictionaries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for real-time compressive sensing in scanning transmission electron microscopy (STEM) face challenges in dictionary learning and reconstruction time, particularly for beam-sensitive materials, limiting the framerate and operational efficiency of live tasks such as focusing and alignment.
Innovation Solution
A method utilizing pre-learned dictionaries to reconstruct electron microscopy images, bypassing the need for extensive dictionary learning by using a 'master' dictionary trained on diverse images, allowing for rapid reconstruction of beam-sensitive materials.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional dictionary learning methods are used for real-time CS-STEM, then reconstruction quality can be achieved, but the computational time and complexity increase significantly, limiting the framerate for live tasks
Solution Approach 1:
The patent applies preliminary action by pre-learning dictionaries from a dataset of fully-sampled STEM images before actual imaging. This pre-learned dictionary is then reused during real-time compressive sensing reconstruction, eliminating the need for time-consuming dictionary learning during live operations. The system performs the computationally intensive dictionary learning phase in advance, allowing rapid reconstruction during actual use.
Solution Approach 2:
The patent extracts and separates the dictionary learning process from the real-time reconstruction process. By taking out the dictionary learning step and performing it offline on a training dataset, the system isolates the computationally heavy operation from the time-critical reconstruction path, enabling faster real-time performance.
2Object-affected harmful factors
If blind inpainting methods are used to avoid needing fully-sampled images, then beam-sensitive materials can be imaged, but the time required to learn a unique dictionary from each subsampled image severely limits the framerate
Solution Approach 1:
The patent applies universality by creating a single pre-learned dictionary that can be reused across multiple different samples and imaging conditions. Instead of learning a unique dictionary for each image (as in blind inpainting), the system uses one universal dictionary trained on diverse STEM images, making it applicable to various beam-sensitive materials without relearning.
Solution Approach 2:
The system performs dictionary learning in advance on a comprehensive training dataset, so that when imaging beam-sensitive materials, the dictionary is already ready for use. This preliminary preparation eliminates the need for time-consuming dictionary learning during actual imaging of sensitive samples.
3Measurement precision
If fully-sampled scans are performed to train dictionaries, then high-quality reconstructions can be achieved, but beam-sensitive materials suffer significant damage
Solution Approach 1:
The patent uses copying by training the dictionary on fully-sampled images from a training dataset (which can be damaged materials) and then applying this learned dictionary to reconstruct images of beam-sensitive materials using only subsampled data. The training copies the learning process from durable samples, preserving the valuable beam-sensitive samples for actual experimentation.
Solution Approach 2:
The system performs the damaging fully-sampled acquisition and dictionary learning in advance on a training dataset, so that when imaging beam-sensitive materials, only low-dose subsampled scans are needed. The preliminary training phase absorbs the sample damage, protecting the actual sensitive samples.
4Object-affected harmful factors
If subsampled scans are performed to reduce electron dose, then beam-sensitive materials can be imaged with less damage, but the reconstruction process takes significant time
Solution Approach 1:
The patent performs the computationally intensive dictionary learning phase in advance before actual imaging. This preliminary preparation creates a ready-to-use dictionary that enables rapid reconstruction of subsampled images during real-time operations, eliminating the time penalty during actual low-dose imaging.
Data Source
AI summary
A method of reconstructing an electron microscopy image of size [Mx N] pixels of a first sample, the method implemented by a computer comprising a processor and a memory, the method comprising:providing a set of pre-learned dictionaries, including a first pre-learned dictionary including a set of p1 atoms;acquiring a sparse set of S acquired sub-images, including a first sub-image of size [a×b] pixels wherein a, b∈[2, min {M,N}], of the first sample; and reconstructing the electron microscopy image of the first sample using the sparse set of S sub-images of the first sample and the set of pre-learned dictionaries.