Ancient coins' surface inspection with web-based neural RTI visualization

The use of neural encodings has the potential to replace the commonly used polynomial fitting in the analysis of artwork surface based on Reflectance Transformation Imaging (RTI), as it has proved to result in more compact encoding with better relight quality, but it is still not widely used due to the lack of efficient implementations available to practitioners. In this work, we describe an optimized system to encode/decode neural relightable images providing interactive visualization in a web interface allowing multi-layer visualization and annotation. To develop it, we performed several experiments testing different decoder architectures and input processing pipelines, evaluating the quality of the results on specific benchmarks to find the optimal tradeoff between relighting quality and efficiency. A specific decoder has been then implemented for the web and integrated into an advanced visualisation tool. The system has been tested for the analysis of a group of ancient Roman bronze coins that present scarce readability and varying levels of preservation and that have been acquired with a multispectral light dome. Their level of corrosion and degradation, which in some cases hinders the recognition of the images, numerals, or text represented on them, makes the system testing particularly challenging and complex. Testing on such a real case scenario, however, enables us to determine the actual improvement that this new RTI visualization tool can offer to numismatists in their ability to identify the coins.

@InProceedings{RTDGPG23a,
  author       = {Righetto, L. and Traviglia, A. and De Bernardin, M. and Gobbetti, E. and Ponchio, F. and Giachetti, A.},
  title        = {Ancient coins' surface inspection with web-based neural RTI visualization},
  booktitle    = {Optics for Arts, Architecture, and Archaeology (O3A)},
  volume       = {12620},
  pages        = {12620:0D},
  year         = {2023},
  publisher    = {SPIE},
  keywords     = {visual computing},
  doi          = {10.1117/12.2674888},
  url          = {https://publications.crs4.it/pubdocs/2023/RTDGPG23a},
}