CapCognition.Net.LicensePlateDetection 2.2.1

License Plate Detection (LPR) for .NET LTS

High-performance license plate recognition for many European countries, including optional vehicle type detection.
Designed for .NET LTS applications with optimized models for different hardware platforms.

Supported Vehicle Types

• Car
• Motorbike
• Tractor
• Truck

Supported Countries

A, B, BG, CH, CZ, D, DK, E, EST, F, FIN, FL, GB, GR, H, HR, I, IRL, L, LT, LV, NL, P, PL, RO, S, SK, SLO

Features

• New multi-model algorithm with recognition rates above 98%.
• Selectable model qualities for performance optimization depending on hardware.
• CUDA acceleration on Windows and Linux for maximum performance.
  Requires CUDA Toolkit and cuDNN.
• Optional vehicle detection can be disabled for higher performance.
• Automatic model download with SHA-256 integrity validation at runtime if not bundled in the deployment.
• SHA-256 validation for local built-in model streams, including plate, text, wide-text, and vehicle models.
• Background pre-download supported (recognition fails if models are not available at runtime).
• Fluent API support (from version 2.1.0).

Temporal License Plate Tracker (new in 2.2.1)

• LicensePlateTemporalTracker stabilizes recognition callbacks in ParkingOfficer, Barrier and CarDetection mode.
• PreferValidated uses country-validated text when available; RawOnly keeps raw OCR text.
• Country-specific formats retain their separator structure during temporal voting: B-W20 and BW-20 stay distinct. No-Country results retain the previous separator-free behavior.
• Optional retained source/plate snapshots use independent size limits and a byte budget. Subscribers own and dispose emitted StableLicensePlateResult instances.

Examples

See .NET LTS examples at CapCognition GitHub.

Documentation

See Documentation for more details.

Model Downloads

Note: The models can only be used together with the CapCognition License Plate Detection library.
Usage in other libraries is not permitted.

• All models can be downloaded from: Download
• Downloaded built-in models are verified against their published SHA-256 checksums before decryption and loading.
• Local models selected through the built-in model-type helpers are validated automatically. Custom streams can supply an expected SHA-256 value through ModelStreamInfo.

Recognition Modes (new in 2.1.0)

The detection can be configured in different modes depending on application requirements:

1. Country-based plate → then vehicle type (vehicle type is detected only if a plate was recognized).
2. Country-based plate only (no vehicle type detection).
3. Vehicle type only.
4. Vehicle type first → then country-based plate (plate search only inside detected vehicles).
5. "No country" plate → then vehicle type (plate detection without country validation).
6. "No country" plate only (no vehicle type detection).
7. Vehicle type first → then "no country" plate (plate search only inside detected vehicles).

The new "No Country" plate detection allows reliable recognition of most license plates, even if the country is not explicitly trained.
Limitation: No country validation of recognized plates.

Platform Support

• Windows x64
• Linux 64-bit
• Linux ARM 64-bit (Raspberry OS)
  • Raspberry OS 64-bit Bullseye (legacy)
  • Raspberry OS 64-bit Bookworm

Only 64-bit operating systems are supported.

Licensing

This library is available as a trial version with a limited number of recognitions per application start.

• After reaching the limit, the application must be restarted to continue recognition.
• The trial version is intended for testing and development purposes only.
• For production use, a license must be purchased.

Release Notes

Version 2.2.1

• Added the temporal license plate tracker with three operating modes, bounded image retention and country-specific separator-aware stabilization.
• Improved compatibility with the updated asynchronous processing lifecycle in CapCognition.Net.Core
• Improved stability during parallel and continuous recognition
• Added automatic SHA-256 integrity validation for downloaded built-in models
• Added SHA-256 integrity validation for local model streams

Version 2.2.0

• .NET 10
• AOT compilation support
• Library initialization simplified
• Continuous capturing and recognition switched to "Pull-Frame" mechanism for simpler and more efficient stream capturing
• License plate decoding has been updated to handle license plates with spaces (encoded as dashes) to ensure unique identification, particularly for German license plates such as H-HA1 (Hanover) and HH-A1 (Hamburg).
• Loading of local models simplified and optimized

Version 2.1.7

• ISO country codes added to the output for better integration and identification of recognized license plates.
• ONNX 1.26.0

Version 2.1.6

• Updated NuGet dependencies
• ONNX 1.24.2

Version 2.1.5

• Swiss license plate validation fixed

Version 2.1.4

• License plate decoding has been updated to handle license plates with spaces (encoded as dashes) to ensure unique identification, particularly for German license plates such as H-HA1 (Hanover) and HH-A1 (Hamburg).

Version 2.1.3

• option switch fixed and optimized

Version 2.1.2

• Updated NuGet dependencies

Version 2.1.1

• options made disposable and some tool functions added

Version 2.1.0

• Added Fluent API
• New recognition modes (country-based, vehicle-based, "no country" just license plate)

Version 2.0.3

• NuGet dependencies updated