A Comparative Study of Pruning Strategies for EfficientNet and ResNet Architectures in Search and Rescue Applications

Abstract

Nigeria's security challenges necessitate effective surveillance solutions. This study evaluates EfficientNet-B0, EfficientNet-B6, and ResNet-18 for Search and Rescue (SAR) operations using computer vision. We applied L1 Unstructured and Random Unstructured pruning techniques to assess each model's accuracy and computational efficiency. Through preprocessing diverse images and testing pre-trained models, our findings reveal that L1 unstructured pruning significantly improves processing times while preserving accuracy. Among the evaluated backbones, EfficientNet-B0 with L1 pruning emerged as the most efficient and accurate for SAR applications. This study offers valuable insights into selecting optimal computer vision models and pruning strategies for enhanced real-world surveillance.