融合语义知识的深度表达学习及在视觉理解中的应用

发布时间：2018-05-05 20:35

本文选题：深度学习 + 神经网络　；参考：《计算机研究与发展》2017年06期

【摘要】：近几年来,随着深度学习技术的日趋完善,传统的计算机视觉任务得到了前所未有的发展.如何将传统视觉研究中的领域知识融入到深度模型中提升深度模型的视觉表达能力,从而应对更为复杂的视觉任务,成为了学术界广泛关注的问题.鉴于此,以融合了语义知识的深度表达学习为主线展开了一系列研究.取得的主要创新成果包括3个方面:1)研究了将单类型的语义信息(类别相似性)融入到深度特征的学习中,提出了嵌入正则化语义关联的深度Hash学习方法,并将其应用于图像的相似性比对与检索问题中,取得了较大的性能提升;2)研究了将多类型信息(多重上下文信息)融入到深度特征的学习中,提出了基于长短期记忆神经网络的场景上下文学习方法,并将其应用于复杂场景的几何属性分析问题中;3)研究了将视觉数据的结构化语义配置融入到深度表达的学习中,提出了融合语法知识的表达学习方法,并将其应用到复杂场景下的通用内容解析问题中.相关的实验结果表明:该方法能有效地对场景的结构化配置进行预测.
[Abstract]:In recent years, with the improvement of deep learning technology, the traditional computer vision task has been developed unprecedented. How to integrate the domain knowledge of traditional visual research into the depth model to enhance the visual expression ability of the depth model, so as to deal with more complex visual tasks, has become a widespread concern in the academic community. In view of this, a series of researches are carried out based on the deep representation learning of semantic knowledge. The main innovation results obtained include three aspects: 1) this paper studies the integration of single-type semantic information (category similarity) into the learning of depth features, and proposes a method of deep Hash learning with embedded regularized semantic association. It is applied to the similarity comparison and retrieval of images, and a great performance improvement is achieved. (2) the multi-type information (multi-context information) is integrated into the depth feature learning. A scene context learning method based on short and long memory neural networks is proposed and applied to the geometric attribute analysis of complex scenes. (3) the structural semantic configuration of visual data is integrated into the learning of deep representation. This paper proposes a representation learning method of fusion grammar knowledge and applies it to the general content resolution problem in complex scenarios. The experimental results show that the proposed method can effectively predict the structured configuration of the scene.
【作者单位】：中山大学数据科学与计算机学院;
【基金】：国家自然科学基金优秀青年科学基金项目(6162200366)~~
【分类号】：TP18;TP391.41

【相似文献】