基于深度卷积神经网络构建场景部件的场景图像分类研究

Scene categorization has a wide range of applications in the areas of multi-media information management, computer vision and intelligent robotics. With the development of information technology, solving the problem of scene categorization on complex scene datasets becomes urgent. Because complex scene datasets are characterized of large numbers of categories to separate, huge inter-class similarity and inner-class variability, it is quite challenging to perform scene categorization on complex scene datasets. To solve this problem, in this project, we propose to create scene part models based on deep convolutional neural networks for scene categorization. The main contents of the research include: transferring visual concepts learned by deep convolutional neural networks for selecting scene parts automatically for scenes in complex scene datasets; creating scene-specific part models based on convolutional neural network features for improving the ability of models for discriminating scenes; integrating appearance information and spatial information of scene parts for improving their ability to discriminate scenes further; combining random forest and support vector machines for creating ensemble classifiers that are adaptable to patterns of hard samples, for classifying image representations effectively. Research of this project is of great significance to the exploration of the problem of scene categorization on complex scene datasets.

场景图像分类在多媒体信息管理、计算机视觉以及智能机器人等领域应用前景广泛。随着信息技术的发展，解决在复杂场景图像集上的图像分类问题变得十分迫切。针对复杂场景图像集因场景类别多，图像存在类内差异大且类间相似度高等特点而难以有效分类的问题，本课题提出基于深度卷积神经网络构建场景部件模型的场景图像分类方法，研究内容包括：针对复杂场景图像集中部件难以确定的问题，通过迁移深度卷积神经网络所学视觉概念，实现部件的自动选取；针对当前场景部件模型分辨力较低的问题，基于深度卷积神经网络特征构建场景类别特有的部件模型，以提高部件模型分辨力；研究整合场景部件的外观信息和空间分布信息的方法，以通过利用两者的相关性，进一步提高部件模型分辨力；通过融合随机森林和支持向量机，构建对困难样本模式具有自适应能力的集成分类器，以对基于部件的图像表示进行有效分类。本项目研究对探索解决复杂场景图像集上的图像分类问题具有重要意义。

场景图像分类在多媒体信息管理、人机交互和智能机器人等领域应用前景广泛。为解决这些领域对复杂场景图像集上的图像分类问题的迫切需求，本课题针对复杂场景图像集因场景类别多，图像存在类内差异大且类间相似度高等特点而难以有效分类的问题，提出基于深度卷积神经网络构建场景部件模型的场景图像分类方法，主要研究成果包括：利用深度卷积神经网络所学的视觉概念及相应视觉概念在目标场景图像集上的响应的统计实现为每个场景类别自动选取相应的场景部件；基于隐支持向量机的学习方法，通过迭代的方式，将选取的场景部件改良为场景类别特有的部件模型，提升了部件模型的区分力；提出了基于长短期记忆深度神经网络协调深度卷积神经网络所提特征的场景图像分类方法，实现了对场景图像的外观信息和空间分布信息的整合；提出融合随机森林和支持向量机的场景图像分类方法，构建对困难样本模式具有自适应能力的集成分类器，对场景图像的表示进行了有效分类。本项目研究对探索解决复杂场景图像集上的图像分类问题提供了新思路和理论依据，并为场景图像分类技术在相关领域的应用提供了技术支持。