Optical Structure Recognition (OSR) — the pipeline that detects molecular structure images on patent and journal pages and converts them into InChI and SMILES strings — has advanced rapidly thanks to modern deep learning, yet its end-to-end accuracy is still bounded by two upstream steps: how precisely the detector localizes each structure, and how cleanly the cropped image is passed to the OCR model. This presentation will share two recent improvements to our internal pipeline. First, the structure detection model has been retrained on the latest YOLO architecture, which removes Non-Maximum Suppression (NMS) and Distribution Focal Loss (DFL) in favor of a leaner inference path; the new weights deliver more precise mask segmentation and tighter bounding boxes, improved recall on structures the previous model missed (including starting compounds in reaction schemes and substances embedded in tables), the elimination of thirty or more Markush fragments that were previously captured as false positives, and meaningfully faster inference — with the maintainers reporting up to 43% gains on CPUs and the path now open to deployment on CPU and edge devices. Second, an optional post-processing step based on Maximum Connected Components (MCC) has been introduced between detection and OCR to clean residual artifacts from the mask-cropped image, with a known caveat for tightly-spaced table layouts where the white space between the structure and surrounding ink is too narrow for the algorithm to safely separate; in those cases, the pipeline still produces a usable OCR output that an analyst can finish cleaning. Both changes are intended to reduce the volume of "junk" that subject-matter experts (SMEs) must clean during curation, with at-scale SME validation as the natural next step.
Video Stream
