A Novel Dynamic Model Switching Approach for Deep Learning on Edge Devices with Performance Evaluation

We propose a novel multi-stage dynamic modelswitching pipeline that enables complex deep learning tasks to run on highly resource-constrained edge devices such as the Arduino Nano 33 Ble Sense and Arduino Nicla Vision. This approach overcomes the RAM, Flash, and compute limitations that prevent the simultaneous execution of multiple vision and audio models on MCUs. Starting with voice recognition, results show that while single-word detection is reliable, there is high misclassification in multiple-word phrase detection, due to the silence between words being interpreted as background or some other class. Vision-based tasks were also explored, where simple image classification and object detection proved to be inadequate for real-world currency recognition. To overcome this, a modular pipeline is implemented: a FOMO-based object detection model first identifies symbolic anchors (Targeting the RBI logo and the Lion Emblem here) of the note, and the cropped region between them is fed as an image to a custom digit classifier model trained to identify numeric digits at a close distance. Memory and hardware constraints prevent simultaneous model execution, so a novel dynamic model-switching mechanism is used to run each model separately on RAM. We achieved 96% accuracy in digit classification and 88.79% accuracy in logo detection model, while running at 66 ms and 65 ms latency.