This phase of the Headstone Photograph Processing System continued the effort to automate the painstaking work of matching photographs of headstones to their official cemetery records for the National Cemetery Administration. Like earlier work on the project, the system takes raw photos of headstones and runs them through a pipeline that detects and straightens each stone, crops it to a clean image, reads the engraved text, and matches that text to the correct record. This 2021 team focused on strengthening the underlying computer vision techniques—experimenting with ResNet-based classification, the EAST text-detection model, and Fourier-transform-based image rotation to better handle the curved, weathered, and irregularly angled text common on real headstones—and refining the fuzzy-matching step that links imperfect text readings to the right database entry. The goal remained the same as the project’s original vision: dramatically reducing the hours of manual cropping, transcription, and record-matching that humanities researchers and cemetery staff would otherwise face when digitizing thousands of veterans’ grave records.
