Various Convolutional Neural Networks for Predicting Types of Hurricane Damage Images

The difficulty of rescue after a hurricane is often reflected in the time delay caused by the inability to accurately detect the damage of the building and the injury caused by the secondary collapse of the building. The existing seismic reconnaissance methods are mainly manual operation. A team of certified inspectors visits the affected building, takes measurements from the damaged area, processes the information gathered, and makes a retrograde decision. However, this procedure is time-consuming and expensive as it requires a lot of manpower. In this experiment, this paper judged the performance of different neural network models e.g. MobileNetV2 and VGG by comparing their accuracy for classifying the same set of images. This study also built a simple new vanilla model for comparison. Among them, the vanilla model achieved 92% accuracy in image classification, which is slightly inferior to the classic model MobileNetV2 +Adam (96.7%), which reflects the possibility of using machine learning neural network to help post-disaster rescue image analysis.