Flood Event Recognition from Images Based on CNN and Semantics
Yanlin Teng, Bo Lang, Zepeng Gu · 2017
Recognizing sensitive events in images, such as flood events, is significant for the maintenance of normal public opinion and social stability. By now, it is still a challenging problem. In this paper, we propose a novel method for recognizing flood events using semantics and CNN-based multi-label image classification. Our method utilizes the empirical information of people so as to improve the recognition result. We first define the semantic model of flood event and introduce it into a multi-label classification model which is based on CNN, and then use the event discrimination model to predict the results. Concretely, for the multi-label classification, we propose a novel end-to-end model, which takes the images as inputs and outputs the predictions of multi-labels. The model can be pre-trained with a single-label dataset, and no ground-truth bounding boxes are required for training. The experimental results demonstrate that our proposed flood event recognition method is superior to others.