基于改进的Faster RCNN面部表情检测算法

PDF下载HTML

Title:

Facial expression recognition based on improved Faster RCNN

作者:

伍锡如, 凌星雨

桂林电子科技大学 电子工程与自动化学院，广西 桂林 541004

Author(s):

WU Xiru, LING Xingyu

College of Electronic Engineering and Automation, Guilin University of Electronic Technology, Guilin 541004, China

关键词:

目标检测; 深度学习; 表情识别; 快速区域卷积神经网络; 特征提取; 分类识别; 多目标识别; 多目标定位

Keywords:

target detection; deep learning; expression recognition; Faster RCNN; feature extraction; classification and recognition; multi-target recognition; multi-target location

分类号:

TP391.4

DOI:

10.11992/tis.201910020

摘要:

针对真实环境下多目标表情分类识别算法准确率低的问题，提出一种基于改进的快速区域卷积神经网络（Faster RCNN）面部表情检测算法。该算法利用二阶检测网络实现表情识别中的多目标识别与定位，使用密集连接模块替代原始的特征提取模块，该模块能够融合多层次特征信息，增加网络深度并避免网络梯度消失。采用柔性非极大抑制（soft-NMS）改进候选框合并策略，设计衰减函数替换传统非极大抑制（NMS）贪心算法，避免相邻或重叠目标漏检，提高网络在多目标情况下的检测准确率。通过构建真实环境下的表情数据集，基于改进的Faster RCNN进行实验测试，在不同场景中能够检测出目标的面部表情，检测准确率相比原始检测模型提高5%，取得较好的检测精度。

Abstract:

To address the problem of the low accuracy rate of the multi-target facial expression classification and recognition algorithm in real environments, in this paper we propose a facial expression detection algorithm based on an improved faster region-based convolutional neural network (RCNN). The proposed algorithm uses a two-stage detection network to accomplish multi-target recognition and location in facial expression recognition. Instead of the original feature extraction module, densely connected convolutional networks are used, which can fuse multi-level feature information, increase network depth, and prevent network gradient disappearance. Soft non-maximum suppression (NMS) is used to improve the candidate-box merging strategy, and the attenuation function is designed to replace the traditional NMS greedy algorithm, thereby preventing the missed detection of adjacent or overlapping targets and improving the detection accuracy of the network under multi-target conditions. Through the construction of an expression data set in a real environment and an experiment based on the improved Faster RCNN, the facial expression of the target was detected in different scenes with a detection accuracy rate 5% higher than that of the original detection model. Therefore, good accuracy is achieved by the proposed algorithm.