表情的表达十分地细腻而复杂，一点点小小的面部动作就会带来丰富的表义，这为对表情建立快速而准确认知带来了一定的困难。对于公安、纪检之类的特殊行业，在讯问过程中会频繁进行讯问对象的情绪认知工作，这无疑会带来比较重的工作负担。如果存在一个能够自动对表情信息进行初步提取与分析的系统，就会在一定程度上减少工作人员的工作量，提高高作效率与工作质量。因此，基于这个工程需求，本文中分别实现了表情识别算法、表情相关的局部人脸提取算法与面部活跃值计算模块。针对讯问过程，各个算法能够分别提供表情识别得到的情绪类、情绪相关的局部面部区域提示热力图与面部活跃值三种不同类型的输出，直观易懂且便于进行统计分析，有较好的实际应用价值。 本文的主要贡献如下：1. 完成了基于2D图像的7种离散表情识别分类算法。表情识别算法由ResNet-50特征提取网络、信息聚合网络以及两个噪声抑制模块共同构成。采用在开源数据集上预训练并在实际应用特化数据集上微调的形式完成模型训练。基于应用场景的实际数据，使用算法初标与人工核查的形式构建了一个小型的表情数据集用于识别算法的微调工作。2. 基于对面部特征的不同分析方法，分别实现了与表情相关的局部面部区域热力图可视化与面部活跃值计算的算法，提供了人在做出特定表情时需要关注的面部区域的热力图提醒以及对象在讯问过程中的细节面部纹理变化的监测。

The expression of expressions is very delicate and complex, which makes it difficult to cognize. For special industries such as those related to public security and discipline inspection, the emotional perception of the interrogation subject will be carried out frequently during the interrogation process, which will undoubtedly bring a relatively heavy workload. If there is a system that can automatically extract and analyze expression information, it will reduce the workload of staff to a certain extent and improve the efficiencyand quality of work.Therefore, based on this engineering requirement, an expression recognition algorithm, an expression-related partial face extraction algorithm, and a facial activity valuecalculation algorithm are respectively implemented in this paper. For the interrogationprocess, each algorithm can provide three different types of output, namely the emotionclass obtained by expression recognition, the emotion-related partial facial area promptheat map, and the facial activity value, which are intuitive, easy to understand, and easyto carry out statistical analysis. The main contributions of this paper are as follows:1. Completed the classification algorithm for 7 kinds of discrete expression recog-nition based on 2D images. The expression recognition algorithm is composedof ResNet-50 feature extraction network, information aggregation network andtwo noise suppression modules. Model training is done in the form of pre-training on open-source datasets and fine-tuning on practical application-specificdatasets.Based on the actual data of the application scenario, a small facial expres-sion dataset was constructed for fine-tuning the recognition algorithm in the formof initial algorithm standardization and manual verification.2. Based on different analysis methods of facial features, the algorithms for visual-ization of heat maps of local facial areas and calculation of facial activity valuesare realized respectively. It provides heat map reminders of facial areas that peo-ple need to pay attention to when making specific expressions, and monitoring ofdetailed facial texture changes of subjects du