Implementation and Performance Evaluation of Convolutional Neural Network models for Low-Power Microcontrollers with Constrained Resources

Recent advancements in machine learning have given rise to TinyML, a field focused on developing efficient, miniature models capable of operating on devices with severe power and computational limitations. In this paper, we evaluate the performance of TensorFlow Lite Micro Convolutional Neural Network (CNN) models, which are prime examples of TinyML. Our research centers on image classification tasks, with a strong emphasis on enabling execution on sensor node devices equipped with ARM Cortex M4 microcontrollers. With a specific focus on the application of TinyML in underwater sensor networks, where resource limitations are paramount, our study serves as a benchmark, assessing the capabilities of these lightweight CNN models across low-power sensor nodes characterized by diverse computational and memory constraints. Our findings convincingly demonstrate the practicality and adaptability of TinyML models on low-power devices based on ARM Cortex M4 microcontrollers. The overarching goal of this research is to contribute to a broader understanding of the potential of TinyML in critical real-world applications, where energy and bandwidth resources are scarce, and the need for immediate data processing is imperative.