Overview

This project implements an end-to-end pipeline to detect and read license plates using OpenCV and Tesseract OCR, as required by the project specifications. It compares performance on Russian (black-and-white) plates versus a non-Russian, multicolored Colorado plate.

Approach

• Plate Localization: Pre-trained Haar-cascade model in OpenCV using the cv2 Python library
• Preprocessing: Grayscale conversion —> adaptive thresholding —> morphological filtering.
• OCR: Tesseract via pytesseract, with confidence-based filtering and conditional preprocessing loops.

Key Outcomes

• Detection
  • 100% success on four Russian plates
  • 0% detection on the Colorado sample
• OCR First Pass
  • One plate yielded the full eight-character string (50% correct)
• Preprocessing Experiments
  • Multiple parameter sweeps increased character regions but did not improve recognition accuracy

Challenges and Learnings

• Plate design variance (shape, color) heavily impacts a fixed Haar cascade.
• Static thresholds struggle with multicolored backgrounds and decorative graphics.
• Rule-based OCR pipelines require extensive tuning and still face recognition limits.

Next Steps for Future Work

• Swap Haar cascade for a modern object detector (e.g., YOLOv5) for robust localization
• Add color-space preprocessing (HSV/Lab) to better isolate text on colorful plates
• Evaluate end-to-end trainable OCR models to boost recognition accuracy

Feel free to explore the full write-up and code on GitHub!