SE-Pruned ResNet-18: Balancing Accuracy and Efficiency for Object Classification on Resource-Constrained Devices

Deep learning-based image object classification methods often achieve high accuracy, but with the growing demand for real-time performance on resource-constrained edge devices, existing approaches face challenges in balancing accuracy, computational complexity, and model size. To alleviate this awkward situation, we propose a novel ResNet-18 architecture that integrates the Squeeze-and-Excitation (SE) module and model pruning. The SE module adaptively emphasizes informative feature channels to enhance classification accuracy, while pruning technology reduces computational costs by removing unimportant connections or parameters without significant accuracy loss. Extensive experiments on benchmark datasets demonstrate that the optimized model outperforms the original ResNet-18 in both accuracy and inference speed. The classification accuracy increases from 93.2% to 94.1%, the number of parameters is reduced by 30%, the Floating-Point Operations decreases from 1.81 giga to 1.32 giga, and the inference time is decreased from 15.2 milliseconds to 12.8 milliseconds per batch. What’s more, the proposed model outperforms MobileNetV2, ShuffleNetV2, and EfficientNet-B0 in accuracy while maintaining competitive inference speed and parameter count. The experimental results highlight the model’s potential for deployment on resource-constrained devices, expanding the practical application scenarios of object classification methods in edge computing and real-time detection tasks.