The integration of deep learning (DL) into Internet of Things (IoT) systems requires careful selection of neural network architectures, deployment frameworks, and model compression strategies because IoT devices usually operate with limited RAM, processing capacity, storage, and battery power. This paper investigates the performance of three DL architectures, namely CNN, LSTM, and CNN-LSTM Hybrid, along with five deployment frameworks: TensorFlow Lite, PyTorch Mobile, ONNX Runtime, OpenVINO, and TVM, across five major IoT application domains. An experimental analysis was conducted on 48 controlled system configurations using standard IoT benchmark datasets. The evaluation metrics included inference latency, classification accuracy, F1-score, energy consumption, peak RAM usage, and throughput. Pearson correlation, one-way ANOVA, and regression modelling were applied for statistical validation, while MobileNetV2 was tested under post-training quantisation from FP32 to INT4. Results show that Hybrid Edge-Cloud achieved the best latency–accuracy balance, with 28 ms latency and 92.1% accuracy. CNN-LSTM Hybrid outperformed standalone models, achieving a mean F1-score of 93.9%. TVM produced the highest inference speed, and INT8 quantisation reduced MobileNetV2 size by 75% with only 0.8% accuracy loss. Therefore, CNN-LSTM Hybrid with INT8 quantisation and TVM optimisation is recommended for practical IoT deployment.
Singla et al. (Tue,) studied this question.