Skip to content
David Lonski
← Back to research

layer: archive · research log

OCR & Object Recognition

Undergraduate research with Dr. Ning Yu · 2025

OCR pipelines for high-resolution image datasets, pairing classical computer vision with the Gemini API for context-aware recognition.

  • Python
  • Gemini API
  • OpenCV
  • NumPy
  • Matplotlib

what I did

  • 01Built a small OCR app: snap a photo, extract the text with Gemini, then run a second Gemini pass to correct it.
  • 02Used region-of-interest boxes to find and scale text inside mixed images: drawings, cars and street signs.
  • 03Located each character in generated CAPTCHA text and returned its coordinates.
  • 04Worked through diverse, edge-case-heavy visual data where off-the-shelf recognition falls short.
  • 05Built the image processing layer with OpenCV, NumPy and Matplotlib.

from the work

The OCR app reading the text on a box of crayons held up to a webcam
The OCR app: Gemini reads the text off a photo, then a second pass corrects it.
Region-of-interest boxes drawn on a drawing of an apple, a car and a street sign
Region-of-interest boxes for finding text in drawings, cars and street signs.
Boxes around each character of a CAPTCHA, with the detected text and coordinates below
Character-level detection on generated CAPTCHA text, with coordinates for each letter.

next log

Mask R-CNN — Image Forgery Detection →