Two-Stream Convolutional Neural Networks for Emergency Recognition in Images

Abstract

Emergencies threaten the safety of public lives and properties. If news agencies can timely report emergencies, their subsequent hazards can be significantly reduced. However, in face of massive pictures, the traditional manual screening can no longer meet the needs for news agencies. Therefore, it is necessary to use a more effective method to classify emergencies, which help news agencies choose the right pictures and release them to the public in time. This paper proposes a method to classify emergencies in still images using two-stream convolutional neural networks(CNNs). Firstly, the architecture of our two- stream CNNs is decomposed into object net and scene net, which extract useful information from the perspective of objects and scene context, respectively. Meanwhile, we investigate different methods for two-stream CNNs feature fusion to improve the performance of emergency recognition. Secondly, another binary classifier works after the two-stream CNNs to verify whether the result of the two-stream CNN is the true positive instance of the predicted emergency class. Experimental results confirm the effectiveness of the proposed emergency recognition method.