基于记忆的不变图像特征学习方法研究

项目名称： 基于记忆的不变图像特征学习方法研究

项目编号： No.61502195

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 其他

项目作者： 魏艳涛

作者单位： 华中师范大学

项目金额： 20万元

中文摘要： 图像特征学习是图像分类的关键环节，是模式识别与计算机视觉领域中的研究热点。传统特征提取方法受到复杂背景、旋转以及尺度变化等多种因素制约，导致图像分类系统的泛化性较差且样本复杂度较高。本项目拟面向图像分类，引入流形学习、不变群以及熵等理论开展基于记忆的不变特征学习方法研究。首先，融合流形学习理论，研究基于记忆的分层学习模型中的编码理论与方法，并利用编码输出的统计特征探讨具有向量形式输出的汇聚方法，提升特征对旋转、尺度变化以及局部形变等变化的不变性，降低图像分类系统的样本复杂度；其次，根据基于记忆的特征学习方法中不同层次的特点研究模板学习理论与方法，进一步增强图像特征的不变性与判别性，降低复杂背景的影响，提高分类精度；最后，从群和熵的角度分析特征学习方法的不变性与判别性，揭示分层不变特征学习方法的内在机制。本研究将进一步丰富图像特征学习的理论基础，为计算机视觉系统的设计和开发提供关键技术支撑。

中文关键词： 图像特征；不变特征学习；分层模型；图像分类；深度学习

英文摘要： Image feature learning is the key part of image classification, and a research hotspot in the fields of pattern recognition and computer vision. However, the traditional feature extraction methods have constraints of complex background, rotation, scale changes, and so on. This makes image classification systems cannot achieve good generalization performance and have high sample complexity. This project will study memory-based invariant feature learning method for image classification by introducing manifold learning, invariant group and entropy theory etc. Firstly, this project will study the coding theory and method in the memory-based hierarchical learning model by fusing manifold learning, and pooling method which makes use of the statistical characteristics of the output of the coding layer and has the outputs with vector form. This study can enhance invariance to rotation, scale changes, local deformation, and so on. And the invariant features can reduce the sample complexity of image classification systems. Secondly, template learning methods in the memory-based invariant feature learning method will be designed according to the characteristic of different layers. The proposed methods can enhance invariance and discriminability of image features, reduce the impact of the complex background, and improve the classification accuracy. Finally, the invariance and discriminant analysis of the proposed feature learning methods will be given from the perspectives of group and entropy. It can reveal the mechanism of the hierarchical invariant feature learning methods. This research will enrich theoretical basis of image feature learning and provide key techniques for the design and develop of the computer vision systems.

英文关键词： Image Feature;Invariant Feature Learning;Hierarchical Model;Image Classification;Deep Learning