openbr简介

OpenBR [3] is a framework for investigating new modalities, improving existing algorithms, interfacing with commercial systems, measuring recognition performance, and deploying automated biometric systems. The project is designed to facilitate rapid algorithm prototyping, and features a mature core framework, flexible plugin system, and support for open and closed source development. Off-the-shelf algorithms are also available for specific modalities including Face Recognition, Age Estimation, and Gender Estimation.

OpenBR originated within The MITRE Corporation from a need to streamline the process of prototyping new algorithms. The project was later published as open source software under the Apache 2 license and is free for academic and commercial use.

Openbr是一个框架，用于调研新模型、提升已有算法、和商业模型组合，衡量识别以及构建自动化的生物识别系统。这个项目用于快速迭代算法模型、构建一个成熟的核心架构、灵活插件系统、支持开源以及封闭的源代码。现成的算法也同样适用于这套框架，包括人脸识别、年龄预估和姿态判断。

Openbr起源于MTTRE组织，一个简化原型设计的新算法的组织。这个项目后来在开源软件公布，并且免费用于学术和商业。

The two principal software artifacts are the shared library 'openbr' and command line application 'br'.

1.架构 

（1）数据结构

在openbr中有两个重要的数据结构和6个接口，数据结构（1）文件结构体：存储文件路径和存储关联元数据的KV表；（2）计量生物结构体：opencv矩阵和关联文件。

（2）插件

在openBR中使用插件是添加新算法的首选方式。

总共有6个抽象接口：实现新文件格式化：Format 、Gallery、Output；实现新的特征计算和比对：Transform、Distance；初始化接口：Initializer。

A Format represents a template on disk either before or after enrollment. For example, images, videos, MATLAB
matricies, and many other extensions can be interpreted by format plugins.
A Gallery represents a template list on disk either be fore or after enrollment. The NIST .xml signature set and OpenBR binary .gal are the standard plugins for storing template lists before and after enrollment, though many others exist including Weka .arff.
An Output represents the result of comparing two galleries. The NIST .mtx binary similarity matrix is the preferred output,though many others exist including .rr rank retrieval and .csv plain text score matrix.
A Transform is a single step in a template generation algorithm, it applies the same image processing or numerical analysis algorithm to every template it receives. Transforms can be either trainable (e.g., LDA) or untrainable (e.g.,LBP). Time-varying transforms also exist to support objecttracking in video.
A Distance is capable of comparing two templates and returning a similarity score. OpenBR supports many common similarity metrics including norm-based, cosine, Chi-squared, and Bhattacharyya. Section 4.5 discusses a
particular distance metric novel to OpenBR.
Commercial algorithms can also be added to OpenBR by wrapping them in Transform and Distance plugins. To date, six commercial systems have been leveraged through the OpenBR API.

2.人脸识别

（1）人脸检测

openBR对opencv的VJ对象识别方法进行了封装，同时提供了基于casecade的正脸识别方法。对于人眼识别，提供了以及基于c++、接口为ASEE的人眼识别方法ASEEEyes。

（2）规范化

人脸基于人脸位置做旋转扭正。

（3）特征描述

基于人脸网格提取LBP和SIFT特征，LBP为基于6*6的窗口每次计算8*8像素，SIF在12像素的半径范围内提取10*10的元素，PCA对每个局域特征进行压缩，生成最终的特征空间，并进行格式化。

（4）特征提取

提取12维特征，每一维是总特征的5%，然后使用LDA方法学习空间特征。

（5）特征比对匹配

计算所有特征间的最大值和最小值

（6）最终算法