Inspiration

I was inspired by the need for privacy, speed, and offline AI capabilities on mobile devices. Arm’s dominance in mobile architecture motivated me to build a solution that leverages on-device AI for real-time image recognition, making advanced vision accessible without relying on the cloud.

What it does

ArmVision performs real-time image recognition directly on Arm-based mobile devices. It classifies images instantly, ensuring privacy, low latency, and efficient battery usage, all without needing an internet connection.

How I built it

Trained a lightweight CNN model for image classification using TensorFlow. Optimized the model with quantization and converted it to TensorFlow Lite for mobile deployment. Integrated Arm NN and Android NN API to maximize hardware acceleration. Built an Android app to capture images and display results in real time. Tested and profiled the app on multiple Arm-based smartphones and tablets.

Challenges I ran into

Balancing model accuracy and size for mobile deployment. Achieving optimal performance across different Arm chipsets. Debugging and optimizing inference speed and battery usage. Ensuring compatibility with various Android versions and hardware.

Accomplishments that I'm proud of

Achieved real-time inference (< 500ms per image) on mid-range Arm devices. Reduced model size to under 5MB while maintaining >90% accuracy. Successfully demonstrated local AI inference, enhancing privacy and efficiency.

What I learned

Practical skills in optimizing neural networks for mobile. Deep understanding of Arm NN, Android NN API, and TensorFlow Lite. Importance of hardware-software co-design for efficient AI on mobile. Effective debugging and profiling techniques for mobile AI applications.

What's next for ArmVision: Real-Time On-Device Image Recognition for Mobile

Future plans include:Expanding to object detection and style transfer. Supporting more device types and Android versions. Integrating additional Arm-specific optimizations for further performance gains. Exploring integration with other mobile AI frameworks and sensors.