Abstract:

With the development of intelligent transportation,it has become an urgent problem to quickly and accurately recognize complex traffic scene.In recent years,a large number of scene recognition methods have been proposed to improve the effectiveness of traffic scene recognition,however,most of these algorithms cannot extract the semantic characteristics of the concept of vision,leading to the low recognition accuracy in traffic scenes.Therefore,a novel traffic scene recognition algorithm which extracts the high-level semantic and structural information for improving the accuracy was proposed.A system to discover semantically meaningful descriptions of the scene classes to reduce the “semantic gap” between the high level and the low-level feature representation was built.Then,the multi-label network was trained by minimizing loss function (namely,element-wise logistic loss) to obtain the high-level semantic representation of traffic scene images.Finally,experiments on four large-scale scene recognition datasets show that the proposed algorithm considerably outperforms other state-of-the-art methods.

Key words: scene recognition, CNN, high-level feature, low-level feature