基于自适应组岭估计的多标签回归面部表情分析方法

21 浏览量更新于2024-08-29 收藏 496KB PDF 举报

"AN ADAPTIVE GROUP LASSO BASED MULTI-LABEL REGRESSION APPROACH FOR FACIAL EXPRESSION ANALYSIS" 这篇研究论文聚焦于面部表情分析，采用了一种自适应组lasso（Adaptive Group Lasso）为基础的多标签回归方法。在传统的面部表情分析中，通常尝试将每一张面部图片与预定义的情感类别关联起来。然而，实际生活中的面部表情往往并不完全符合这些固定的情感状态。因此，为了更准确地解析面部表情，论文提出了一种新的方法，该方法用多个预定义情感状态的连续值来描绘每个面部表情，从而打破了单一标签的限制。自适应组lasso被用来描述不同标签之间的关系，即不同的面部表情在某些情感相关的面部区域（patches）上共享相似性。这种方法能够捕捉到不同情感状态之间的复杂联系，并且能够在具有共同特征的面部区域上进行有效的特征选择和压缩。为了解决多标签回归问题，论文提出了一个凸优化公式化方法。这种公式化方式确保了能够找到全局最优解，从而提高模型的预测能力和泛化能力。通过在标准的JAFFE数据集上进行实验，结果验证了所提方法的优越性能。该实验可能涉及了准确率、召回率、F1分数等评估指标，以展示提出的多标签回归方法在面部表情识别上的优势。关键词：面部表情分析，自适应组lasso，多标签回归，凸优化，情感状态，JAFFE数据集。

AN ADAPTIVE GROUP LASSO BASED MULTI-LABEL REGRESSION APPROACH

FOR FACIAL EXPRESSION ANALYSIS

Kaili Zhao, Honggang Zhang, Jun Guo

Beijing University of Posts and Telecommunications, China

ABSTRACT

In the realm of facial expression analysis, numerous attempt-

s have been made to link each facial picture to one affective

category. Nevertheless, in our daily life, few of the facial ex-

pressions are exactly one of the predeﬁned affective states.

Therefore, to analyze the facial expressions more effectively,

this paper proposes an Adaptive Group Lasso based Multi-

label Regression approach, which depicts each facial expres-

sion with multiple continuous values of predeﬁned affective

states. Adaptive Group Lasso is adopted to depict the rela-

tionship between different labels which different facial ex-

pressions share some same affective facial areas (patches).

Moreover, to solve the multi-label regression problem, a con-

vex optimization formulation is presented, which would guar-

antee a global optimal solution. The experiment results based

on JAFFE dataset have veriﬁed the superior performance of

our approach.

Index Terms— Facial Expression Analysis, Multi-label

Regression, Adaptive Group Lasso

1. INTRODUCTION

Nowadays, facial expression analysis techniques are becom-

ing increasingly important in our daily life. Most previous at-

tempts depict each facial picture with one of the predeﬁned af-

fective labels, such as the 6 basic affective states of happiness,

sadness, surprise, fear, anger and disgust [1, 2]. Numerous

recently researches have been focused on feature extraction

methods [3, 4] or classiﬁers [5, 6] and satisﬁed single-label

recognition accuracy has already been obtained.

However, one common concern of the previous approach-

es is the assumption that each facial picture is linked to only

one affective label, which tends to be over simpliﬁed. As we

know, one can produce complex expressions. In 1971, Ekman

and Friesen [7] describe which blends of the basic emotions

occur and what these blends look like universally. For in-

stance, when one gets an unexpected award, such as the ICIP

Best Student Paper, he would be both surprised and happy.

JAFFE dataset [8], which is one the most widely used dataset

in the realm of facial expression analysis, is originally de-

scribed via multi-label, that is the continuous values of the

6 basic affective states, which represents the mean semantic

ratings of each picture by 60 Japanese female students. Mean-

while, no pictures in the dataset can be lossless described by

one of the predeﬁned affective states

Therefore, to depict the facial expression more effective-

ly, this paper proposes an Adaptive Group Lasso (AGL) based

multi-label regression approach for facial expression analysis.

With each picture describing by continuous values of all the

6 affective states, we obtain a far more accurate way to repre-

sent facial expressions. AGL is adopted to depict the relation-

ship between different labels: different affective states share

nearly the same facial affective areas, which is the facial parts

to express emotions. In the past decades, there are many stud-

ies focusing on facial expression analysis with Action Units

which depicts facial muscle distributions [2, 9]. This makes

us consider an interesting problem how to utilize facial affec-

tive areas to analyze facial expressions. For example, though

happiness and sadness are totally different affective states, the

same areas of the face, such as mouth, eyebrow and jaw, are

the facial parts to express both emotions. Thus, by placing

the weights, which are extracted from the same position for

different labels, into the same group, the AGL regularization

term in the objective function would encourage them to be ze-

ro or non-zero simultaneously. Therefore, the nonzero-weight

areas of different labels would turn to be the same. Mean-

while, by solving the multi-label regression problem via a

convex optimization formulation, our AGL method enjoys a

global optimal solution during the training process.

To evaluate the performance of our multi-label regression

algorithm, we adopt the original multi-label data of JAFFE

and compare our algorithm against the start-of-the-art algo-

rithm of Support Vector Regeression (SVR), as well as the

baseline algorithm of Linear Regression (LR). According to

the evaluation criterions of Sum Absolute Error (SAE) and

Sum Squared Error (SSE), our algorithm obtains superior per-

formance.

The remainder of the paper is organized as follows. To

start with, Section 2 illustrates that facial expressions recog-

nition should be modeled as a multi-label regression problem.

In Section 3, AGL and its relationship with facial expression

analysis is described. Then, Section 4 introduces our multi-

label regression algorithm based on AGL. In Section 5, the

experiment results are presented and Section 6 draws the con-

clusion.

下载后可阅读完整内容，剩余4页未读，立即下载

weixin_38704835

粉丝: 4
资源: 936

基于自适应组岭估计的多标签回归面部表情分析方法

部分线性模型的adaptive group lasso变量选择研究

数字图像处理课程：adaptive group lasso在变量选择中的应用

数字图像处理： Adaptive Group Lasso在变量选择中的应用

matlab图像相嵌代码-Adaptive-Graph-Guided-Embedding-for-Multi-label-Annotation

An Adaptive Memetic Framework for Multi-objective Combinatorial Optimization Problems: Studies on Software Next Release and Travelling Salesman Problems

DL-Adaptive-Resource-Allocation-for-a-Multi-user-_通讯编程_matlab_

An Adaptive Efficiency-Fairness Meta-scheduler for Data-Intensive Computing

Spatially Adaptive Block-Based Super-Resolution

An improved self-calibration approach based on adaptive genetic algorithm for position-based visual servo (2008年)

Self-Adaptive-Synthetic-Over-Sampling-Approach:题为“一种用于不平衡分类的自适应合成过采样技术”的研究论文的 MATLAB 代码。-matlab开发

最新资源