GeoDict-AI
The GeoDict-AI module covers the functionalities in GeoDict that aim to reconstruct 3D models obtained from segmented computer tomography or FIB/SEM images of different multi-component materials, e.g. nonwovens and electrodes.
The GeoDict-AI module provides the possibility of using Artificial Intelligence (AI) approaches for the separation of the different material while image segmentation.
Often, different materials, for example binder and fibers, have the same gray values in the images. Therefore, an automatic separation based on the gray value is not possible. However, it is possible to separate the different materials based on their shape in the image. A neural network can be trained to identify binder in a grain structure, to differentiate two fiber types with different shapes or even to identify the individual fibers. In the graphic it is shown how binder is identified from fibers and fibers from each other.
GeoDict-AI is the starting point to analyze physical properties on the segmented 3D-scans considering the different materials. GeoDict-AI can be used for nearly every structure that can be generated with the GeoDict structure generator modules, e.g. grain structures with binder in GrainGeo or fiber structures with binder in FiberGeo. Of course, you can also use other training material. The training data then must be organized in a defined folder path.
Neural networks can be trained for four different application cases. In the following for each of them a recommended workflow is outlined.
Identify individual fibers
To train a neural network to identify individual fibers, the following workflow can be applied:
- Start with creating a generation script (.py) using the GeoDict structure generator modules. Select the right parameter ranges to match the statistical properties for the considered fibers that should be identified, e.g. fiber diameters, orientations, and the structure density.
- Then, use the script to design your experiment and to create the training and testing data.
- Having the training data, train a neural network that can identify the individual fibers in every segmented 3D-scan containing fibers matching the statistical properties caught in the training data.
- Validate the performance of the trained neural network by applying it to all generated training and testing structures.
- Finally, use the model in FiberFind to identify individual fibers in a segmented 3D-scan.
|
Distinguish between different materials
For identifying different materials, the following workflow can be applied:
- Start with creating a generation script (.py) using the GeoDict structure generator modules. Select the right parameter ranges to match the statistical properties for the considered materials that should be distinguished, e.g. the object diameters, the object orientations, and the solid volume percentages.
- Then, use the script to design your experiment and to create the training and testing data.
- Having the training data, train a neural network that can identify the target materials in every segmented 3D-scan containing materials matching the statistical properties caught in the training data.
- Validate the performance of the trained neural network by applying it to all generated training and testing structures.
- Finally, use Apply Neural Network to apply the trained network to segmented 3D-scans and identify the desired target material.
|
Enhance grayscale images
Another application of AI is enhancing gray value images. For example, to reduce time and costs in generating CT-scans, a neural network can be trained to enhance the image quality. For example, a scan can be taken from only a few angles or with reduced exposure time.
- Create low-quality / high-quality image pairs with the same resolution following certain requirements.
- Set-up the training data following a defined folder path.
- Using the training data, a neural network can be trained.
- Apply the network on scans with Enhance Grayscale Image.
|
Segment grayscale images
Since GeoDict 2025 it is also possible to train a neural network to Segment Grayscale Image.
- Generate training data. One possibility is to generate training data using the GeoDict structure generators and transform the structures to gray value images using the GeoApp Generate Artificial CT Scan. The other possibility is to use real scans as described here.
- Set-up the training data following a defined folder path.
- Given your training data, a neural network can be trained.
- Apply the neural network with Segment Grayscale Image.
|
