The fundamental idea of AI is that a neural network is trained to perform a special task, such as the identification of separate fibers, the differentiation between material phases, or to enhance or segment a grayscale image. To train the neural network, enough examples need to be available to teach it. Clearly, for 3D-scans it is a very hard and time-consuming task to manually label materials or objects for the number of examples required. This is where GeoDict’s unique structure-generation capabilities come in. The idea is that a neural network trained on these synthetic samples can then be used to analyze 3D scans of real materials. E.g.:

• FiberGeo can quickly generate a large number of 3D structures of fibrous media. Variation of fiber diameters, shapes, lengths, curvatures and density as well as the amount of binder, if any, can be specified to match the ranges seen in the real materials to be analyzed.
• GrainGeo provides similar possibilities to generate different grain structures.

As long as these structures are close enough to the real 3D-scans, they can be used to train neural networks using GeoDict-AI. For Enhance Grayscale Image, you should provide pairs of registered low-quality / high-quality (CT) scans. The neural network then learns to transform low-quality images to high-quality images. Note however that even a small number of image pairs (e.g. 2-3) can be sufficient, depending on the application. Depending on the scan, it can also be possible to create such images using the GeoApp Generate Artificial CT-Scan. This app is used in combination with the GeoDict structure generators, generating an image based on a structure file. Thus, this app can also be used to create training data to segment grayscale images.

When applied to a 3D image, the neural network will transform it to another image by assigning a scalar output value between 0 and 1 to each voxel. How this image is interpreted depends on the task at hand:

• When identifying fibers, the neural network will mark a centerline which corresponds to a curve along the central axis of each fiber. FiberFind-AI will then automatically reconstruct the individual fiber objects from this output.
• When identifying material phases (e.g. binder in a fibrous structure) the different material classes are marked directly by the neural network output.
• For enhancing grayscale images, the neural network output corresponds directly to the improved grayscale image.
• When segmenting grayscale images a input grayscale image is transformed into a segmented structure file.

For a detailed explanation of neural network training look here.

GeoDict-AI uses the Pytorch Framework by The Linux Foundation, one of the most used and well-known machine learning libraries (see also Paszke et. al., 2019). The required version of Pytorch is installed during the GeoDict installation and GeoDict-AI works out of the box. Install GeoDict on the used machine, to install Pytorch correctly.

GeoDict-AI may run on the CPU (the main processor) or on the GPU (the graphics card). If one or more suitable GPUs are detected during installation of GeoDict, GeoDict-AI will run on the GPU; otherwise it will run on the CPU.

The GPU version is usually much faster than the CPU version (roughly a factor 10), but it requires a good GPU. For up-to-date recommendations visit our website.