Efficient Device-Edge Inference for Disaster Classification

Image classification can learn useful insights from crisis incidents and is gaining popularity in the field of disaster management. This is fueled by the recent advances in computer vision and deep learning techniques, where accurate neural network models for disaster type classification can be accrued. However, these studies quite commonly neglect the prohibitive inference workload which may hamper its wide-spread deployment, especially for model execution on low-powered edge devices. In this paper, we propose a lightweight disaster classification model that recognizes four types of natural disaster plus one non-disaster class. To support real-time applications, the proposed model is optimized with OpenVINO, which is a neural network acceleration platform. Different from existing works which focus on benchmarking at training stage, our experimental results reveal the actual performance at inference stage. Specifically, the optimized version achieves up to 23.93 frames per second (FPS), which is more than doubled the speed achieved by the original model, while sacrificing only 0.935 % of classification accuracy.