License Plate Recognition using Computer Vision and OCR

Project Goals

• Detect license plates across different framing, lighting, and background conditions.
• Extract alphanumeric plate text from detected regions with a practical OCR workflow.
• Build a modular pipeline that separates detection, cropping, preprocessing, and recognition.
• Use saved result files to review confidence, exact matches, and failure cases honestly.

Data and Inputs

The project uses 433 labeled vehicle images from public license-plate datasets, plus saved model outputs in CSV form. The source material includes sample inputs, preprocessing-output visuals, confidence summaries, and prediction-review tables.

The page intentionally excludes private keys and vendor-specific notebook configuration. The portfolio focuses on the pipeline, outputs, and evaluation framing.

Pipeline

The workflow separates detection, cropping, image preprocessing, OCR, and review. Fourier, soft-threshold, and binary preprocessing variants were compared so OCR quality could be debugged in context rather than treated as one opaque end-to-end result.

This structure makes the system easier to improve because a bad read can be traced to localization, contrast, preprocessing, character recognition, or post-processing.

Evaluation and Output Review

The saved analysis file includes 360 rows with final predictions. Among 331 rows with both a normalized label and a normalized prediction, 150 were exact OCR matches. I present that as an audit result rather than production accuracy because dataset quality, plate validity, and OCR post-processing all affect the number.

The most useful project evidence is the error analysis: high confidence did not always mean a correct full plate, which is a realistic ALPR failure mode.

Limitations and Next Steps

• Plates vary in size, angle, contrast, background, and formatting.
• OCR accuracy drops quickly when detection crops are noisy or low contrast.
• End-to-end performance depends on the interaction between detection and text extraction.
• A stronger next version would add plate-format validation, character-level edit distance, more robust crop filtering, and separate detection metrics.

What It Shows

This is a strong applied computer-vision case study because it deals with an imperfect real-world problem rather than idealized input data.

The project demonstrates practical judgment around localization, preprocessing, OCR, and model-evaluation tradeoffs.