Tiny-EfficientNet-SE: A Squeeze-and-Excitation Enhanced CNN Architecture for Multi-Domain Classification on Edge Device

The growing demand for efficient and compact deep learning (DL) models designed for resource-constrained edge applications has surged with the increasing deployment of edge intelligence solutions. For image recognition tasks, lightweight Convolutional Neural Network (CNN) models have been developed for real-time and edge deployment. This paper presents Tiny-EfficientNet-SE model, an enhanced lightweight CNN architecture designed specifically for edge computing applications. This architecture integrates Squeeze-and-Excitation (SE) Blocks for adaptive feature recalibration, Batch Normalization (BN) for accelerated convergence, and applies knowledge distillation techniques from EfficientE0 model for enhanced feature generalization. This architecture addresses the critical tradeoff between high accuracy and minimal resource consumption with just 0.46 MB model size, while achieving accuracies of 93%, 95%, 97% and 98% for multi-class classification across human face emotion, cereal crop pest, pneumonia disease, and Weather condition datasets, respectively. Compared with Latest lightweight CNN models such as MobileNet, SqueezeNet, EfficientNet, and ShuffleNet, Tiny-EfficientNet-SE model exhibited superior performance in model size and comparable accuracy, making it highly suitable for deployment on edge and resourceconstrained environments, as well as advancing real-time edge intelligence solutions.