基于耦合边缘化自动编码器的跨域行人重识别
基于耦合边缘化自动编码器的跨域行人重识别(任务书,开题报告,论文10500字)
摘要
行人重识别，旨在从不同地点配置的各种相机识别同一行人的图像。运用计算机视觉、机器学习方法进行行人重识别可以有效帮助工作人员在海量视频中快速发现、追踪行人目标，本文主要针对行人图像跨视角且分辨率不同这一具体问题，应用耦合边缘化自动编码器方法来实现行人重识别。
耦合边缘化自动编码器模型构建了两个边缘化去噪自动编码器分别用于低清图像、高清图像的重构，同时，为了耦合了两个自动编码器的学习添加了一个特征映射在高清域和低清域之间迁移知识，最后在低清重构的输出层上学习行人间的区分性特征，进行行人重识别。本文的主要工作如下：
针对模型构建，在已有的耦合边缘化自动编码器中，使用了高清矩阵和低清矩阵的依赖关系，重新建立了模型的迭代优化学习方案，同时引入低清重构的权重，使模型更有利于解决行人重识别问题。
在基于patch特征的模型构建中应用显著性方法进一步优化性能，并将优化后的patch模型和基于image特征构建的模型进行融合。实验结果表明，融合后的模型行人匹配性能得到比较明显的改进，在VIPER数据集上， rank1达到了 22.78%，取得了较好的结果。 [资料来源：http://Doc163.com]
基于Java Swing实现了行人匹配的可视化展示，对于每个待查询行人，我们将展示检索库中与其最相近的前21个行人图像，同时标记正确的行人图像。
关键词：行人重识别；跨域多视角；耦合边缘化自动编码器
Abstract
Person re-identification, aimed at recognizing images of same person from various cameras at different locations. The use of computer vision and machine learning methods for person re-identification can effectively help workers quickly find and track target person in massive videos, such as lost children, the elderly, and suspects.This dissertation focuses on the specific issue of cross-view and different resolutions of pedestrian images. We use Coupled Marginalized Auto-Encoders for cross-domain multi-view learning.
Coupled Marginalized Auto-Encoders designs two Marginalized Auto-Encoders for the reconstruction of low-resolution person images and high-resolution person images. To better couple the two denoising auto-encoderslearning, they incorporate a feature mapping, whichtends to transfer knowledge between the high-resolution domain and low-resolution domain. Finally, they learn more discriminative features on the output layer of low-resolution domain to solve person re-identification problem. The main work is as follows: [资料来源：Doc163.com]
For the model construction, based on the existing Coupled Marginalized Auto-Encoders, the dependence relationship between the high-resolution matrix and the low-resolution matrix is used to re-establish the model's iterative optimization learning scheme. At the same time, the weights of the low-resolution reconstruction are introduced, and the model is more conducive to achieving pedestrian fine-grained classification.
In addition, we apply salience methods on the patch-based model for further performance optimization. The optimized patch model and image model are merged to get the final score of each pair of person. The experimental results show that the merged model's pedestrian matching performance has been improved significantly.On the VIPER dataset,Our Rank1 reached 22.78%, and achieved good results.
Finally, we made a visual display of the matching results based on Java Swing.For each pedestrian to be queried, we will display the first 21 pedestrian images closest to it in gallery and mark the correct pedestrian image. [资料来源：http://doc163.com]
Key Words:Person re-identification; cross-domain multi-view; coupled marginalized automatic encoder;
[资料来源：http://Doc163.com]
目录
摘要 I
Abstract II
第 1 章绪论 1
1.1 背景 1
1.2 目的及意义 2
1.3 国内外研究现状 2
1.4 课题研究内容 3
第 2 章针对行人重识别问题的耦合边缘化自动编码器框架 5
2.1 耦合边缘化自动编码器 5
2.2 针对跨域多视角行人重识别问题的模型 6
2.3 模型迭代优化学习方案 10 [资料来源：www.doc163.com]
2.4 本章小结 13
第 3 章模型构建 14
3.1 基于patch特征的自动编码器模型构建 14
3.2 基于patch模型的无监督显著性优化方案 16
3.3 基于image特征的自动编码器模型构建 18
3.4 模型融合及结果分析 18
3.5 基于Java Swing的匹配结果可视化 21
3.6 本章小结 22
第 4 章总结与展望 23
4.1 总结 23
4.2 展望 23
参考文献 25
致谢 27