Tiny-MobileNet-SE: A Hybrid Lightweight CNN Architecture for Resource-Constrained IoT Devices
Traditional Convolutional Neural Network (CNN) architectures face challenges in deployment on resource-constrained devices such as Internet of Things (IoT) platforms, mobile applications, and drones due to their computational intensity and memory requirements. This limitation motivates the developme...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11048474/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Traditional Convolutional Neural Network (CNN) architectures face challenges in deployment on resource-constrained devices such as Internet of Things (IoT) platforms, mobile applications, and drones due to their computational intensity and memory requirements. This limitation motivates the development of lightweight models that maintain high performance with minimal resource consumption. This paper introduces Tiny-MobileNet-SE, a novel hybrid CNN architecture designed for image classification tasks under critical resource-constrained environments. This architecture integrates Squeeze-and-Excitation (SE) blocks for adaptive feature recalibration, Batch Normalization (BN) for accelerated convergence, and applies knowledge distillation techniques from MobileNetV2 for enhanced feature generalization. We conducted comprehensive experimentation to validate its performance across various metrics, including accuracy, number of parameters, and model size. Experimental results show that Tiny-MobileNet-SE achieved superior performance with a significantly reduced model size (0.72 MB) and a smaller number of parameters (482 K) compared to conventional lightweight CNNs, reaching up to 98.6% validation accuracy on real-world crop pest and disease datasets. This innovative architecture contributes to bridging the gap between computational efficiency and classification accuracy, enabling deep learning (DL) models to be deployed on small mobile, embedded, and IoT devices with minimized resource consumption. |
|---|---|
| ISSN: | 2169-3536 |