Inspiration Laboratory sample labels contain critical information, but handwriting, glare, and curved tubes can make them difficult to read. Inspired by the MRG Laboratories challenge, we built LRS-3 — Label Reading System to reduce repetitive transcription and help operators turn sample labels into organized records.

Our idea was simple: Three tubes. One scan. Clear sample records. What it does LRS-3 combines a rotating three-tube holder with camera capture and label recognition. Its automatic capture workflow collects eight views per tube, producing 24 individual tube images grouped into three separate samples.

The scanner decodes barcodes, extracts label fields, and displays readings alongside their image crops for human review. Operators can correct uncertain values, confirm records, and export a CSV.

An xAI Grok assistant explains OCR warnings, suggests better photography techniques, and—with explicit permission—offers a second reading of difficult fields. Operators remain responsible for confirming the results.

An earlier saved-photo benchmark averaged approximately 20.8 seconds of processing per tube, excluding capture, upload, and human review. We are targeting 90% field-reading accuracy, which still requires evaluation on unseen samples. How we built it We manually photographed 40+ physical samples, collecting 150+ images for model training and development. Multiple views helped us capture different sides of curved labels and variations in handwriting, lighting, and camera angle.

We connected this data work with a physical scanning prototype:

  • Hardware: We designed parts in OpenSCAD and printed them using a Bambu P1P. The assembly includes separate gears, washers, and holders for approximately 17 mm tube caps.
  • Motion: An Arduino UNO Q controls a 28BYJ-48 stepper motor through a ULN2003 driver. The gear system rotates all three tubes together.
  • Capture: A browser camera interface coordinates stationary photographs with motor movements and settling pauses. The automatic sequence uses seven movements between eight capture positions.
  • Processing: A Python/FastAPI backend combines OpenCV, ZXing barcode decoding, RapidOCR, and local recognizers.
  • Storage and review: SQLite tracks samples and readings. The interface supports corrections, confirmation, and CSV export.
  • Generative AI: The xAI Grok API provides optional assistance for difficult label readings while keeping suggestions separate from confirmed records. Challenges we ran into Small printed locking parts lifted from the build plate, forcing us to revise their geometry and adhesion settings. We also had to balance rotation speed with reliable motor movement and stable photography.

Connecting an iPhone camera introduced challenges with network isolation, changing laptop addresses, HTTPS certificates, camera permissions, and mobile usability.

The labels themselves were another challenge. Handwriting, reflections, and curved surfaces made a single photograph insufficient in some cases. These difficulties reinforced the importance of multiple views, preserved image evidence, and human review. Accomplishments that we're proud of We brought mechanical design, motor control, label recognition, and AI-assisted review into one prototype.

Our key milestones include:

  • Manually collecting 150+ images from 40+ physical samples for training and development.
  • Building a geared fixture that rotates three tubes together while keeping their captured images grouped separately.
  • Implementing an automatic workflow for eight views per tube.
  • Completing a motor movement-and-settling test and confirming smooth physical rotation.
  • Reaching approximately 20.8 seconds of average processing per tube in an earlier saved-photo benchmark.
  • Passing 71 Python tests, plus JavaScript checks for image separation, capture sequencing, and failure handling. We also built a Grok review workflow that helps operators investigate uncertainty without automatically accepting AI-generated values. What we learned Reliable scanning depends on more than an OCR model. Lighting, focus, mechanical alignment, and capture timing all influence the quality of the final reading.

Collecting our own dataset taught us how much effort goes into useful training images. Multiple photographs provide more evidence, but they do not replace having diverse physical samples.

We also learned that commanded motor steps do not guarantee physical position, and that separate sample records require careful image grouping. Most importantly, we learned to treat AI output as a suggestion that needs supporting evidence. What's next for LRS-3 Our next milestone is a complete end-to-end phone scan with the assembled hardware. We plan to improve mobile usability, simplify installation on another laptop, and refine the camera mount and lighting.

We will evaluate unseen labels against independently transcribed answers and measure progress toward our 90% field-reading accuracy target. All photos of a physical sample will remain in the same dataset split to prevent training and evaluation overlap.

We also plan to measure the full capture-to-export time and determine whether Grok assistance reduces review time without increasing accepted mistakes.

Built With

Share this project:

Updates

Submission history