Limited Resources Usage of Deep-Learning-Based IoT Devices in Healthcare

Deep learning models are widely used in healthcare. Given the large amount of data and the need for real-time processing, running these models on IoT devices is a viable solution. This study examines the impact of deep learning models on the limited resources of IoT devices. We selected three models introduced at the 2023 CVPR conference and executed them on a Raspberry Pi with different dataset sizes to measure CPU utilization, memory utilization, and energy consumption. The results show that for the pruned Yolov5 model, CPU utilization has the strongest positive correlation with energy consumption ($5.43 \times 10=4$ Joules), while memory utilization shows the most significant negative correlation ($-4.27 \times 10=4$ Joules). In the FasterNet-T0 model, dataset size has the strongest positive correlation with energy consumption ($2.75 \times 10=5$ Joules). For the Lite-Mono model, dataset size shows the most significant positive correlation ($1.84 \times 10=4$ Joules), while memory utilization has the strongest negative correlation with energy consumption (- $2.52 \times 10=4$ Joules). These results highlight the key factors correlated with energy consumption across different models, providing valuable insights for deploying deep learning models on IoT devices.