The innovative nature of this study lies in precisely quantifying the loss of classification accuracy resulting from the transfer of a special AI model for species classification from data obtained using a high-resolution CT device to data obtained using a lower-resolution device, and in proposing strategies for closing this performance gap.
The investigations for this initial study were carried out on the basis of pine and maple samples, while the high-resolution scanner was a sub-µCT device with a resolution range of several hundred nanometers and the lower-resolution scanner was a µCT device with a resolution range of several µm.
The results of these experiments may contribute in the future to the operation of AI-based wood classification systems with a heterogeneous volumetric CT scanner portfolio.
The scientists hope that the investigations will provide information about the transferability of AI models between different imaging modalities, which benefits from learning robust, wood species-specific patterns from the volumetric image data.
Contact:
More informations:
Link to the publication on researchgate.net
