Abstract:

With the rapid development of deep learning, the scene recognition method based on deep learning gradually replaces the traditional scene recognition method based on manual features, and becomes the main research direction in the future. According to the method of scene recognition based on deep learning, this paper summarizes the basic ideas and divides it into four categories： the method of scene recognition based on deep learning and bag of visual words, the method of scene recognition based on salient part, the method of scene recognition based on multi-level feature fusion and the method of scene recognition based on knowledge representation. The characteristics and limitations of each method are analyzed, and the recognition effect is compared. Finally, the future research direction is prospected.

Key words: scene recognition, deep learning, bag-of-visual-words, salient object, multi-layer feature fusion, semantic relation

摘要：

随着深度学习的快速发展，基于深度学习的场景识别方法逐渐取代传统的基于手工特征的场景识别方法，成为未来研究的主要方向。针对基于深度学习的场景识别方法，对基本思想进行了总结，将其大体分为以下四类：深度学习与视觉词袋结合场景识别法、基于显著部分的场景识别法、多层特征融合场景识别法、融合知识表示的场景识别法，分析了各个方法的特点及局限性，并对识别效果进行了比较，最后对未来研究方向进行展望。

关键词: 场景识别, 深度学习, 视觉词袋, 显著目标, 多层特征融合, 语义关系