Problem

The problem was to develop an image classifier using the TensorFlow framework and implement AI inference on an STM32 microcontroller. The challenge was to optimize the model for deployment on a resource-constrained device like STM32, which has limited memory and processing power.

We had to ensure that the model's accuracy was not compromised while reducing its size and complexity. Additionally, we had to integrate the TensorFlow-Lite library with STM32CUBIDE, a development environment for STM32 microcontrollers.

Click here to see a video of our project in action.

Solution

To address these challenges, we first developed an MNIST image classifier using the TensorFlow framework. We then optimized the model by reducing its size and complexity without compromising its accuracy. We used techniques like pruning, quantization, and weight sharing to achieve this.

Next, we integrated TensorFlow-Lite with STM32CUBIDE and implemented AI inference on an STM32 microcontroller. We used CMSIS-NN (Neural Network kernels for Cortex-M) library provided by ARM to accelerate neural network computations on Cortex-M processors. Finally, we tested our solution by deploying it on an STM32F746G Discovery board.

Impact

Reduced model size by X%
Inference time reduced by X%
This project demonstrated how AI can be deployed on resource-constrained devices like microcontrollers without compromising performance. It has potential applications in various fields like robotics, IoT, and edge computing.

Team

Team Member #1 - Machine Learning Engineer
- Developed the image classifier using TensorFlow and optimized the model for deployment on STM32.
- Integrated TensorFlow-Lite with STM32CUBIDE and implemented AI inference on an STM32 microcontroller.
Team Member #2 - Embedded Systems Engineer
- Provided technical expertise in developing and testing the solution on an STM32F746G Discovery board.
- Optimized neural network computations using CMSIS-NN library provided by ARM.

AI Image Classifier on STM32 Microcontroller