Highly Efficient Vision AI for Low Power Microcontrollers
How Focoos and STMicroelectronics are Enabling Real-Time Vision AI on Low Power Edge Hardware
STMicroelectronics
Semantic Segmentation (Pascal VOC)
Object Detection (COCO)
Edge AI & Embedded Vision
Automated model design with Neural Architecture Search
Hardware-aware deployment on STM32N6
Scalable model optimization for edge vision applications
Efficient vision AI on microcontrollers,
without manual tuning.
Overview
Focoos AI and STMicroelectronics worked together to demonstrate that advanced computer vision models can run efficiently on edge-constrained microcontrollers.
The work focused on the STM32N6, which integrates ST’s Neural-ART IP, enabling hardware-accelerated execution of neural networks at the edge.
The shared goal was to make complex Vision AI tasks, such as segmentation and object detection, practical on constrained devices, without requiring manual neural network design or tuning.
Challenge
Deploying modern computer vision models on edge hardware is typically a slow and manual process.
Developers often need to adapt architectures by hand, compress models, adjust parameters, and repeatedly test performance, especially given limited memory, performance, and power.
To make Vision AI scalable on microcontrollers, the process needed to become fully automated, hardware-aware, and efficient, while still delivering real-time performance.
Approach
We used ANYMA, Focoos AI’s Neural Architecture Search (NAS) engine, to automatically generate, evaluate, and deploy neural network architectures optimized specifically for the STM32N6.
More than 2,000 architectures were generated and benchmarked using ST EdgeAI Core Technology, enabling the system to identify the models that are best supported by:
– the STM32N6 hardware pipeline
– the Neural-ART IP compiler toolchain
This removed the need for handcrafted architectures, iterative tuning, or ML domain expertise.
Results
The resulting models run natively on the STM32N6 with ST’s Neural-ART IP and ST EdgeAI Core, with no manual adjustments required.
They deliver real-time performance, with higher accuracy and speed compared to existing edge-focused vision solutions, while operating efficiently within the compute and memory profile of a microcontroller.
Semantic Segmentation
on Pascal-VOC
(21 classes, 512p resolution)
760 models were ranked and successfully deployed on the STM32N6 with ST’s Neural-ART IP.
Object Detection
on COCO
(80 classes, 512p resolution)
1.340 models were ranked and successfully deployed on the STM32N6 with ST’s Neural-ART IP.
“The cooperation with Focoos AI was eye opening for me. The Anyma tool enabled the fully automated design, exploration, and deployment of several thousand complex edge AI topologies on the cutting-edge STM32N6 platform, which integrates ST’s internally developed Neural-ART IP, all without requiring advanced machine learning expertise. Anyma is a highly valuable tool, especially for many small and medium enterprises across the world.”
Danilo Pau – Technical Director, IEEE, AAIA and ST Fellow