Announcement: ESPHome Native Component

[nliaudat]

Hello to all users of the fantastic AI-on-the-edge-device project!

After three months of intensive development, I'm thrilled to announce a new, native ESPHome component that replicates the core digit recognition features: esphome_ai_component.

GitHub Repository: https://github.com/nliaudat/esphome_ai_component

This component is built from the ground up for ESPHome, offering a streamlined, efficient, and deeply integrated experience.
Why Choose This Component?

• Massively Simplified Setup: This is a single YAML configuration in your ESPHome device. Flash it and you're done!

• Blazing Fast & Memory Efficient: The core inference and image processing pipeline for an 8-digit meter takes less than 3 seconds on an ESP32-CAM. It's optimized to run smoothly within the constrained resources of an ESP module.

• Fully Native ESPHome Integration: Get your meter readings directly as ESPHome sensors. Use them natively in Home Assistant automations, dashboards, and with its entire ecosystem (history, graphs, alerts) without any custom MQTT or REST API bridges.

• Months of Improvements: This isn't just a port; it's an evolution. Key enhancements include:

  • Advanced Pre-processing: Incorporates sophisticated image preparation (contrast enhancement, noise reduction, sharpening, and custom windowing/zoom for OV2640 seriescameras) directly on the ESP32, ensuring high recognition accuracy.

  • Optimized TensorFlow Lite Model: Features a highly efficient model with on-demand loading of TFLite operators, reducing the required Tensor Arena memory by half compared to standard setups.

  • Unmatched Hardware & Ecosystem Integration:

  • No More Board Integration Worries: It works on all 180+ officially supported ESPHome microcontrollers. (including ESP32 family, ESP8266, RP2040 (RP2), BK72xx, RTL87xx, and NRF52) [chatgpt says 641 boards, but unconfirmed]

  • Access the Entire Ecosystem: Seamlessly combine your meter reading with any of the 240+ available ESPHome components (sensors, displays, output switches, etc.) for a truly unified device.

The Numbers Speak for Themselves

To quantify the "less complicated" claim, let's look at the code footprint:

AI-on-the-edge-device (C++ inference core + web server): ~ 8,000+ lines of C++

esphome_ai_component: ~ 1,200 lines of C++

RAM:   [==        ]  16.0% (used 52508 bytes from 327680 bytes)
Flash: [==        ]  19.4% (used 1578846 bytes from 8126464 bytes)

That's an 85% reduction in code complexity! This directly translates to easier maintenance, a smaller attack surface, and a more reliable, focused application.

Who Is This For?

This component is perfect for you if:

• You want a "set it and forget it" meter reading solution.

• You are already using ESPHome and want to keep your ecosystem unified.

• You were intimidated by the complexity of the original project's setup.

• You value speed, low resource usage, and maximum compatibility.

Get Started Today!

The repository includes detailed documentation, a ready-to-use example YAML configuration, and instructions for getting started.

Visit the GitHub page to begin: https://github.com/nliaudat/esphome_ai_component

A huge thank you to the original AI-on-the-edge-device project for the inspiration and proof-of-concept. This new component stands on the shoulders of that giant, aiming to make the technology accessible to an even wider audience.

I welcome any feedback, issues, or stars on the repository! Let's make smart meter reading simpler for everyone.

Footnote:

This project is still in alpha development as it has so far only been tested by me. Feel free to add suggestions and recommendations. If you like the project, please make a PR!

This component focuses on the core digit recognition. Features like the detailed web-based user interface for manual correction from the original project are not part of this native ESPHome implementation.

It has no sdcard support cause I feel it's useless, it can be integrated later

Actually, it's main focused on hassio, but can work later standalone.

I may works for any tflite model, but it needs to be modularized into smaller parts

[caco3]

Looks interesting, but please at leaste note in your repo where you have the models from.

I use ESPHome for several other devices, so this could also be interesting for me.
How ever, as I understand it, as for now it just digitizes, right?
A core part of our project is the whole validation and error handling. I would not want to have to do this in HA...

[nliaudat]

The model source and license is in the ## license section of readme : The AI models from haverland is under Apache Licence

I mean it can works on analog model out of the box by creating a second instance of meter_reader_tflite with analog model. I have no analog meter to test.

The error checking is done in https://github.com/nliaudat/esphome_ai_component/blob/main/components/meter_reader_tflite/output_validation.cpp and can be enhanced

[moritzj29]

Is it possible to also read analog meters with this? The haverland model is only working on digits if see it correctly. Would it be possible to use the analog models as described in the documentation (https://jomjol.github.io/AI-on-the-edge-device-docs/Learn-models-with-your-own-images/). Due to board options I would like to go with the esphome based version preferably…

[nliaudat]

In theory it should works, but I cannot test for now

[b2un0]

thank you for this. migrated my gasmeter already to this one. works like a charm!