# face_expression_recognition_for_elderly

**Repository Path**: qdu2016203500/face_expression_recognition_for_elderly

## Basic Information

- **Project Name**: face_expression_recognition_for_elderly
- **Description**: 面向养老院人脸表情识别算法研究

当前项目在修改中，部分内容未添加引用后期会逐步修改完善。
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2023-03-06
- **Last Updated**: 2023-03-06

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README


面向养老院人脸表情识别算法的研究与应用

摘  要
本文为了在养老院的看护设备或是监控摄像设备中使用表情识别的功能，对老年人人脸表情识别算法作了研究，用来反馈老年人每天的表情变化，评估老年人日常生活下的心理状况，在突发疾病时能作出预警。为老年人看护提供智能化保障。
本文研究目的是针对养老院场景中使用的表情识别算法进行研究，提高对老年人表情识别的准确率。为满足养老院需要通过视频图像进行疾病和心理反馈的需求，本文进行了人脸表情识别算法模型的研究，基于主流的分类任务模型进行改进。首先，由于摄像设备问题导致图片模糊、人脸太小等问题，本文作了研究并对人脸图片数据预处理，经过标注后构建老年人表情数据集。其次，相比起年轻人，老年人脸可用于识别的表情特征不充分，这导致表情标注过程中产生大量噪声标签，进而影响到表情识别模型的训练，为了避免模型过度受到噪声数据的影响，提高模型抵抗标签噪声的能力，本文设计信息互补Mask用来使网络关注到人脸不同区域，降低了网络仅仅关注局部噪声特征的可能性，使网络关注到整体特征。同时，因为模型关注到的多个局部特征存在重合的可能，所以本文设计Mask不相交的操作，让Mask代表的局部特征信息不重合，以此实现特征信息最大化。最后，考虑到只通过老年人人脸某一部分特征去识别表情是不准确的，因此让模型综合多个局部特征进行表情分类，通过用局部特征进行监督训练，以保证充分利用多部分特征，实现对老年人表情识别准确率的提升。
通过调整合适的训练参数后，本文完成表情识别模型的训练，并在公开数据集、老年人表情数据集上测试。本文得到了各个数据集上的识别准确率，其中，在公开数据集RAF-DB上可以达到85%以上的准确率，并且在FERPlus实现的效果也很好。针对自建老年人表情数据集，实现80%以上的准确率。同时，本文验证了噪声数据集上的模型性能，在RAF-DB数据集噪声大量增加时的时候，模型依然能保持较高的识别准确率，证明了模型的鲁棒性。最后，各个实验表明，本文通过实现人脸表情识别算法，能帮助识别老年人人脸表情，以视频技术帮助养老院实现对老年人的疾病预警和心理反馈。

关键词： 深度学习，人脸表情识别算法，老年人人脸数据集， 注意力机制


Abstract
In order to use the facial expression recognition function in nursing home equipment or monitoring camera equipment, this paper studies the facial expression recognition algorithm of the elderly, which is used to feedback the daily expression changes of the elderly and evaluate the psychological status of the elderly in daily life , can give early warning in case of sudden illness. Provide intelligent protection for elderly care.
The purpose of this research is to study the expression recognition algorithm used in nursing home scenes, and improve the accuracy of expression recognition for the elderly. In response to the needs of nursing homes for disease and psychological feedback through expressions, this paper designs and implements a facial expression recognition algorithm model, which is improved based on the mainstream classification task model. First of all, due to problems such as blurred pictures caused by camera equipment and too small and unclear faces, this paper has done relevant research and preprocessed the face picture data, and then constructed a dataset of facial expressions for the elderly after annotation. Secondly, compared with young people, elderly faces have insufficient expression features for recognition, which leads to a large number of noise labels in the expression labeling process, which in turn affects the training of the expression recognition model. In order to avoid the model being excessively affected by noise data, To improve the model's ability to resist label noise, this paper designs an information complementary Mask to make the network focus on different areas of the face, reducing the possibility that the network only focuses on local noise features, and making the network focus on the overall features. At the same time, because the multiple local features that the model pays attention to may overlap, so this paper designs a Mask disjoint operation, so that the local feature information represented by the Mask does not overlap, so as to maximize the feature information. Finally, considering that it is inaccurate to recognize expressions only by a certain part of the face features of the elderly, this paper makes the model integrate multiple local features for expression classification, and supervise training with local features to ensure full use of multi-part features. Improve the accuracy of facial expression recognition for the elderly.
After adjusting the appropriate training parameters, this paper completes the training of the facial expression recognition model, and tests it on the public dataset and the elderly facial expressions dataset. This paper obtains the recognition accuracy on each data set, among which, the accuracy rate can reach more than 85% on the public data set RAF-DB, and the effect of FERPlus is also very good. For the self-built elderly expression data set, the accuracy rate is more than 80%. At the same time, this paper verifies the performance of the model on the noise dataset. When the noise of the RAF-DB dataset increases a lot, the model can still maintain a high recognition accuracy, which proves the robustness of the model. Finally, various experiments show that this paper can help recognize the facial expressions of the elderly by implementing the facial expression recognition algorithm, and use video technology to help nursing homes realize disease early warning and psychological feedback for the elderly.

Key words: Deep learning; facial expression recognition algorithm; elderly face dataset; Attention mechanism