基于卷积神经网络和语义信息的场景分类  

Scene Classification With CNN and Semantic Information

作　　者：张晓明 尹鸿峰[2] 

机构地区：[1]北京交通大学计算机与信息技术学院,北京100044 [2]北京交通大学海滨学院计算机科学系,河北黄骅061199

出　　处：《软件》2018年第1期29-34,共6页Software

基　　金：河北省高等教育科技研究重点项目(ZD2017304)

摘　　要：作为计算机视觉领域的一个重要分支,场景分类在许多场所有着重要和广泛的应用。由于场景的整体布局、目标物体的出现以及位置关系的变化等因素使得整体场景更加多样化,然而,利用场景与目标间的语义信息可以有效克服这个问题。本文提出一种同时利用卷积神经网络与语义信息来进行场景分类的方法,该方法首先利用卷积神经网络获取颜色信息对场景进行初始分类,然后利用图像目标物体与场景间的语义信息对分类结果进行修正。我们与目前流行的场景分类方法进行了实验对比,实验结果表明与利用海量训练数据集得到好的分类效果的方法相比,本文方法利用有限的训练数据集的情况下,仍能取得良好的分类效果。As an important branch of computer vision, scene classification has a wide range of applications. Due to the large diversity of the scene images, scene-level tasks are challenging. However, semantic information is a great help to overcome this problem instead of massive training data. In this paper, we introduce a joint scene classi-fication method simultaneously using convolutional neural network and semantic information. The method firstly initializes the scene type by color features extracted by the neural network, and then corrects the classification results with the semantic contextual information. We apply our method to an indoor scene dataset and compare with the state-of-the-art methods. Experiments show that with a small amount of training images, the joint method achieves the same good effect with that based on massive training data.

关 键 词：计算机视觉 场景分类 语义信息 卷积神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]