Low data volume image transmission

Thanks to seeed studio for developing the Grove Vision AI Module V2. It makes deploying edge AI devices faster. You can find more information about the module here.

The project I originally wanted to complete was face image transmission. Since the model porting has not been completed yet, I will tested the transmission architecture first.

Summary

Using the AI ​​detection model, the camera can reduce the amount of data transmitted, but retain the function of monitoring cats or dogs.

Hardware

Grove Vision AI Module V2

Seeed Studio XIAO ESP32C3

Software

Arduino

Python

HiveMQ

Build

step.1 Upload AI model

Upload AI model using SenseCraft AI. You can follow the getting start.

Here I chose the pet detection model.

Browser searche for cat pictures and tests AI performance.

step.2 combine XIAO ESP32C3 and Grove Vision AI Module V2

Use Aruduino IDE upload Demo 1 sketch to XIAO ESP32C3 and get AI inference results via UART.

searche cat pictures again.

(still working below)

step.3 image transmission

When power on or each 10 min, XIAO ESP32C3 will get capture image from Grove Vision AI Module V2 and send it to MQTT server. The rest of the time, inference is done and the results are awaited. If a cat is detected, the image location data will be sent to MQTT.

step.4 image regeneration

Future work:

"Low data volume face image transmission"

Idea

Although the total population now exceeds 8 billion, it is only 2 to the 33rd power. Assuming that each person has 1, 000 different looks after dressing up, then only 35 bits are needed to number all the different looks of everyone. In this case, when the camera transmits character recognition on the screen, it actually only needs to transmit less than 64bit information.

Of course, it is impossible to handle this in reality, so I want to use current AI generation technology to achieve similar effects.

I will test the feasibility of this method on the face first. Just like the game I played when I was a child, I would cut a magazine model picture into a short section and paste it on my notebook. When I turn the page, it will turn into a different character.

Datasets:

First I had to have a picture of the face, and I searched for the facial recognition data set. The picture below is a comparison chart provided by the Webface260m website, which saved me the work of searching for the datasets.

Use case

Although this project can reduce transmission traffic, it will increase a lot of calculations. The usage scenarios I can think of include space communications and deep-sea wireless transmission or polar camp. In these applications, transmission volume is more expensive than energy. However, face recognition is not very useful in these two scenes. After all, there are not many people in these places.