Deep Inference Optimization for Real-Time Edge Deployment on MicrocontrollersNikodinovska, Angela (Avtor) Perš, Janez (Mentor) Koporec, Gregor (Komentor) edge AIimage classificationneural networksdeep learningmicrocontrollersIn recent years, Artificial Intelligence (AI) has made remarkable advancements, revolutionizing various industries by enabling intelligent decision-making and automation. The rapid growth of AI applications has led to an increased demand for real-time inference. Our aim has been to implement edge AI technology by investigating the feasibility of deploying our existing AI cloud-model on a microcontroller, enabling real-time inference. Moving the AI model from the cloud to the edge offers several advantages, including reduced latency, improved privacy, enhanced reliability, and cost-effectiveness. However, migrating the current AI model to a microcontroller presented technical challenges due to the limited computational resources and memory constraints of such devices. Conventional optimization techniques like pruning and weight clustering were inadequate for microcontrollers. We found that quantization significantly reduced the model size, making it a viable solution. Our secondary objective has been to explore alternative AI models more compatible with edge computing. We tested various models, including MobileNetV2, EfficientNetB0, ResNet50 and CNN-LSTM and developed a novel approach to decrease model size while maintaining accuracy. To ensure a cost-effective solution, we investigated different microcontrollers suitable for edge AI and selected the STM32H747 and ESP32-S3 for evaluation.20242024-08-28 14:25:01Magistrsko delo/naloga160437VisID: 62705COBISS_ID: 208926979sl