Library / SDK
parkpow/deep-license-plate-recognition avatar
parkpow/deep-license-plate-recognition

A plate recognition client library that ships the blurring tool in the same folder

Automatic License Plate Recognition (ALPR) or Automatic Number Plate Recognition (ANPR) software that works with any camera.

672 stars135 forksC++MIT

At a glance

What is it?
Example clients and operational utilities for a commercial recognition service, with the models deliberately absent. The repository holds a redaction tool next to the recognition tools, a camera transfer script, a gate relay controller and an FTP poller that can delete remote files.
Who is it for?
Decide what you are building before you read anything else here. This repository is a set of clients for a hosted recognition service, and the page is explicit that the recognition models are not in it, which means every image you feed these scripts leaves your network for someone else's inference service unless you run the self-hosted SDK yourself.
Can I use it commercially?
Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
Is it still maintained?
Yes. The repository last received commits 72 days ago.
What is it written in?
Mainly C++, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • parkpow/deep-license-plate-recognition672 ★ · +3
Star history
Repository2026-09-112026-11-15
parkpow/deep-license-plate-recognition669672

3 Stars in the last 65 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

The repository says what it is not in the first paragraph

The opening sentence is the one that matters. This repository contains example clients, integrations and operational utilities for a commercial service, and it does not contain the recognition models themselves. Recognition happens either against that service's API or against a self-hosted SDK you run yourself, and the client decides which by a flag. The manifests reinforce that this is a workspace rather than a product: the project file is named `platerec`, versioned `0.1.0`, and carries an empty description, and the tool configuration marks it as not a package at all, so the environment exists for development and nothing is installed from it. The repository has no releases either, so the intended consumption model is cloning the repository and reading the section you need.

The redaction tool sits beside the recognition tools

The second tool in the repository map is the one worth reading first, because it is the one that argues against the others. A redaction script detects plates, including small or barely readable ones, and can save a blurred copy of an image. Its options include splitting a high-resolution image into pieces, which the page says costs three API lookups, leaving plates matching a regular expression unblurred, and ignoring results that have no vehicle bounding box. A vehicle number plate is personal data in most jurisdictions, and a script that finds plates and writes a blurred copy is the correct direction for handling images you do not control. Having it in the same tree as the recognition, transfer and gate-control tools is at least an acknowledgement that both operations exist.

The redaction default skips detections with no bounding box

That third option deserves more than a mention. Ignoring results that have no vehicle bounding box is a sensible way to filter false positives, and it is also the filter that leaves a recognised plate unblurred when the service returns the text without a box around the vehicle. The recognition response the page documents contains both fields, a box and a plate string, so the two can be absent independently. A redaction pipeline that defaults to skipping boxless results therefore has a failure mode where a plate is detected and the image is still written out unblurred, and the page does not say what happens in that case. The documented invocation is short:

bash
python number_plate_redaction.py --api-key MY_API_KEY vehicle.jpg --save-blurred

Anyone using the tool on images they did not take should test that path deliberately rather than assume the detection implies the blur.

Credentials travel as command line arguments, and one option deletes files

Every documented invocation passes secrets in the arguments: an API key flag, a service URL flag, and for the file transfer tool a hostname, a username, a password or a private key. That is the clearest possible interface for a reader and the least safe one at rest, since process arguments are visible to anything on the machine and land in shell history. The page does say to replace the placeholder key locally and not to commit the token, and the repository carries a gitignore and a secrets baseline file alongside a pre-commit configuration, so some hygiene is in place. The file transfer script is the sharpest edge: its delete option removes processed files from the remote server, and the page's advice is to test without it first.

Three dependency contracts, and they do not agree

The root project file lists five runtime libraries, including an exact pin on the imaging library, plus a development group with a formatter, a linter, a property-based testing library, a debugger and a test runner. The quickstart tells you to install two packages with no version constraint at all. The file transfer tool adds a third library, and the directory watcher adds two more, none of which is in the root list. So an exact pin and two unpinned installs of the same library can coexist on one machine, and the page resolves this by saying outright that each subproject has its own dependencies and that you should follow its readme rather than assume the image-client setup covers the repository. That warning is the most useful line in the file.

Two scores in the sample response and one of them is unexplained

The documented response is a list of objects, each carrying a version, a filename and a result list, and each result carries a box with corner coordinates, a plate string and two score fields. The example values are a plate reading, a score just above nought point nine, and a second score just above nought point nine-two. Nothing on the page says what the second score measures, whether it relates to the crop, the vehicle detection or something else, and the page defers the full schema to the hosted API reference rather than restating it. That matters for anyone building a threshold on the output, because a field with no documented meaning is a field nobody should threshold on.

Twelve subprojects and a closing sentence that stops mid-word

The map lists twelve destinations: image recognition, redaction, the file transfer processor, a directory watcher, a stream directory for video, webhooks, a camera and parking-system integration, a blur directory and a video editor, on-premise install tooling, benchmarks, a gate relay controller and examples in three languages. Two Android examples live in other repositories, one owned by the same organisation and one by someone else, and a Node-RED integration lives in a third. The repository root carries tooling for a language that appears nowhere in the map, plus a Windows batch file, a pytest configuration and a lock file. And the last sentence of the visible page breaks off after asking readers starting from another language to convert something.

Editorial conclusion

Decide what you are building before you read anything else here. This repository is a set of clients for a hosted recognition service, and the page is explicit that the recognition models are not in it, which means every image you feed these scripts leaves your network for someone else's inference service unless you run the self-hosted SDK yourself. The plumbing around it is the part that turns a lookup into a pipeline: a watcher that recognises new files and moves them to an archive, an FTP and SFTP poller that can delete processed remote files, a webhook integration, and a gate controller that opens a barrier. Read together with the redaction tool in the same tree, the repository is honest about both directions, and the redaction tool is the one most worth understanding first, because its default skips any detection that has no vehicle bounding box. Before deploying any of it, check what your jurisdiction and your property owners require about retaining plate data, prefer the self-hosted SDK for anything involving cameras you do not own, and pass credentials through the environment rather than the command line, because every documented example puts a token, a password or a service URL directly into the process arguments.

Frequently asked questions

What is parkpow/deep-license-plate-recognition?

A collection of example clients, integrations and operational utilities for a commercial plate recognition service, with separate subprojects for image recognition, redaction, file transfer over FTP or SFTP, a directory watcher, streaming video, webhooks, camera and parking integrations, blur and video editing, on-premise install tools, benchmarks, a gate controller, and examples in C++, C# and Java. The page states plainly that it does not contain the recognition models.

Does the repository include a tool to blur number plates?

Yes. The redaction script detects plates, including small or barely readable ones, and can save a blurred copy of an image. Its options include splitting a high-resolution image into pieces at a cost of three API lookups, leaving plates matching a regular expression unblurred, and ignoring results that carry no vehicle bounding box.

What do I need to run the image recognition client?

Python 3.10 or newer, plus either a hosted API token or a running self-hosted SDK. You clone the repository, create a virtual environment, activate it and install two packages, requests and the imaging library, with no version constraint. The page is explicit that other subprojects carry their own dependencies and that the root environment, managed with a separate tool, does not cover them.

Does the FTP processor delete files from the server?

It has an option that removes processed files from the remote server, and the page advises testing without it first. Authentication can use a password or a private key. Note that the password, the API token and the service URL are all passed as command line arguments in the documented examples, so they are visible in the process list and in shell history.

Official sources

  1. Issues
  2. License: MIT
  3. parkpow/deep-license-plate-recognition on GitHub
  4. Project website
  5. README
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/parkpow-deep-license-plate-recognition.svg)](https://hysenlabs.com/projects/parkpow-deep-license-plate-recognition)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/parkpow-deep-license-plate-recognition.svg)](https://hysenlabs.com/projects/parkpow-deep-license-plate-recognition)

Was this useful?