Introduction

(1)问题描述:

super resolution(SP)问题:Gallery是 high resolution(HR),Probe是 low resolution(LR)。

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

(2)当前存在的问题:

① 当前的半耦合(semi-coupled)矩阵学习是解决SR复原,而不是直接进行行人重识别;

② 行人图片存在噪声,直接使用半耦合矩阵学习无法很好的刻画特征空间。

(3)Contribution:

① 提出一个新的半耦合低秩判别矩阵学习方法(semi-coupled low-rank discriminant dictionary learning approach,SLDL),该方法从图像特征中学习得到高低分辨率字典对,将低分辨率特征映射到高分辨率特征;

② 提出一个多视角 SLDL 方法,对不同类别的特征学习出不同的特征对。

Brief Review

(1)SR问题中的耦合字典训练:

目标函数:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

其中 x 和 y 为HR和LR的一对,且论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP,γ 是平衡因子,D 和 D 为耦合字典,K 为原子数量,N 为训练样本数量,a 为编码系数。

(2)行人重识别问题中的半监督耦合字典学习(SSCDL):

假定 x = {x, x, ..., x}, y = {y, y, ..., y},目标函数:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

其中 α,α 为无标签样本的系数矩阵,α 为带标签样本的共享系数矩阵。

(3)半耦合矩阵学习(SCDL):

目标函数:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

其中 γ、λ、λ、λ 为平衡因子,d,d 为字典矩阵的原子,Λ、Λ 为编码系数矩阵,W 为映射矩阵。

Semi-coupled low-rank discriminant dictionary learning(SLDL)

(1)问题定义:

C 表示相机A拍摄的HR行人图像,C 表示相机B拍摄的LR行人图像,并生成相机A的LR图像 C',即和 C 含有相同的分辨率。

在实际中,低分辨率在不同区域的影响是不一样的,如在单一色调的区域影响较小,在复杂纹理的区域影响较大,因此可以考虑将图像划分为若干patch,并对patch进行聚类,对每个聚类学习一个子字典(sub-dictionary)和映射函数。文章对 C' 和 C 进行划分patch,并使用K-means聚类,在依据 C' 的聚类结果,将 C 划分的patch聚入其中。

定义第 i 个聚类的HR、LR的字典为 D、D, 第 i 个聚类的映射矩阵为 V,获得字典集合有:D = [D, D, ..., D],D = [D, D, ..., D],映射矩阵集合有 V = {V, V, ..., V},c 为聚类数量.

其它参数定义:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

(2)方法概述:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

目标函数:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

其中的约束条件为:学习的字典能够非常好的表示训练样本的内在特征,E,E,E表示噪声。

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

其中:

映射项:论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

子字典表征能力项:论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

(为了让第 i 个子字典对除了第 i 个聚类外的特征,具有较差的表征能力)

低秩正则化项:论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

正则化项:论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

区分度项:论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

(其中 S 和 D 表示正样本对和负样本对,论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP表示 C第 p 个图片的重构特征,论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP表示 C 第q 个图片的重构特征,n 为每张图片的patch数)

(3)优化算法:

① 固定 D、D、V,更新 A、A'、B:

初始化字典和映射矩阵:字典采用PCA偏差进行初始化,映射矩阵采用单位矩阵初始化;

编码系数计算如下,其中d()表示为区分度项的计算函数:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

解法:

对于第一个优化目标,转化为如下问题:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

其中论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

可以转为增广拉格朗日乘子问题(Augmented Lagrange Multiplier problem)【传送门】,即:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

【具体解法在该论文的引用[47][50]中介绍,尚未阅读】

第二第三个优化目标同理。

② 固定 D、D、A、A’、B、V(其中 j ≠ i),更新 D 和 D:

由于受到约束条件论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP等影响,论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP也要进行更新。定义:论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP,D 和 D 的更新过程如下:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

其中:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

且:论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

解法:

将 D 的目标函数更新为:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

其中:论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

转化为拉格朗日乘子问题:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

D 的解法同理。

③ 固定 D、D、A、A'、B、V(j ≠ i),更新 V:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

其中 论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

令:论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP,则目标函数转化为:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

迭代更新如下:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

上式是一个岭回归问题,求解为:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

下式求解为:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

(4)算法总结:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

Multi-view SLDL(MVSLDL)

通过实验发现,不同的特征提取对于高低分辨率的映射效果是不同的。

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

MVSLDL 针对不同类型的特征学习不同的映射矩阵,即:对于每个patch,都提取 HSV、LAB、LBP特征,并学习对应的映射矩阵。

变量定义更新:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

目标函数更新:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

采用上述方法解出每种特征的字典和映射矩阵。

Super-resolution person re-identification with learning dictionaries and mappings

(1)采用SLDL的行人重识别:

① 将Probe中LR图像映射到HR特征中:

定义 y 为第 i 个patch的特征:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

聚类索引 j 通过以下函数计算:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

将特征 yi 映射到HR特征中:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

② 重构Gallery中的图像的特征:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

③ 采用欧氏距离计算,对Probe图像与Grallery图像进行匹配。

(2)采用 MVSLDL 进行行人重识别:(主要流程同上,在重构特征部分改进如下)

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

Experimental results

(1)实验设置:

① 数据集:VIPeR、PRID、CUHK01、i-LIDS、ETHZ;

② 特征提取:HSV、LAB、LBP;

③ 参数设置:λ = 1,λ = 1,λ = 1,λ = 1,

对于 VIPeR,论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

对于 i-LIDS,论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

对于 PRID,论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

对于 ETHZ,论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

对于 CUHK01,论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

设置聚类的数量为64,图片的patch分割为 8*8,每个子字典的原子数量为 48.

(2)实验结果:

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

论文阅读笔记(十九)【ITIP2017】:Super-Resolution Person Re-Identification With Semi-Coupled Low-Rank Discriminant Dictionary Learning-LMLPHP

05-11 11:23