Enhancing Recognition Accuracy of Urban Flooding by Generating Synthetic Samples Using CycleGAN

In traditional supervised learning technique, when the number of label samples is sufficient, the accuracy of the model will be high. But, some problems will happen for this kind of training model to be applied to outdoor environment. First, there are not stable and sufficient training samples for the target model. Second, the labeling cost of samples is always high for retraining a new model after a period of time. In this paper, we design a real-time urban flooding detection system based on CCTV images. In order to solve the problem of insufficient samples for special class such as flooding, we use CycleGAN to generate some synthetic samples for training the model. In order to get good samples from the generated samples, we propose a filtering approach to filter out some bad samples. We take 100 CCTVS in our experiments, and the results show we can enhance the accuracy to 97.2%.