Microcontroller Image Classification
Project Goal: Deploy a machine learning model on a microcontroller to optimize inference speed, performance accuracy, and power consumption.
Outcome: A convolutional neural network model predicts landscape images from six classes with 80% accuracy and with a time for inference of 41.1ms (~24.32 inferences/second). Model runs on STM32 Cortex M4 Core.
Poster download: [PDF]
Key Features:
Machine learning model:
- Built and trained a convolutional neural network (CNN) using TensorFlow, outputting landscape classification
- Tested model architectures for optimal performance of 80% accuracy with minimal Flash and RAM usage
Hardware analysis:
- Quantized and pruned the model to reduce the size and memory usage for efficient inference
- Model size of 563 kB (71% of board Flash) and RAM requirement of 68 kB (55% of board SRAM)
Optimization:
- Used a Prune Quantize-Aware Trained (PQAT) model for Flash reduction of 3x while maintaining <1% performance loss
Images:
