Abstract: To address the issues that micro-expression sub-pixel deformations are easily overwhelmed by global noise, standard Transformers suffer from attention dilution under small-sample conditions, and deep features tend to become homogenized, this study proposes a hierarchical Transformer-based micro-expression recognition method integrating dynamic token pruning and hierarchical feature reconstruction. In the preprocessing stage, a motion magnification and Region-based Cosine-weighted Fusion strategy was introduced, employing cosine similarity as a gating coefficient to adaptively suppress the weights of noise regions irrelevant to facial expressions. In the feature encoding stage, a three-layer heterogeneous Transformer architecture was designed, sequentially embedding a dynamic token pruning module to discard low-quality tokens, a mild compression strategy to model cross-regional muscle synergies, and a hierarchical feature recombination module to break deep-layer homogenization. In the aggregation stage, a convolutional block attention module was adopted to calibrate and enhance discriminative motion responses. On the composite MEGC 2019 dataset, the proposed method achieves UF1 and UAR scores of 89.10% and 88.73%, respectively. On the CASME II, SAMM, and SMIC subsets, it attains optimal UF1 scores of 96.16%, 85.26%, and 83.39%, demonstrating the effectiveness of the proposed architecture.

Key words: Micro-expression recognition, Optical flow partitioning, Hierarchical Transformer, Deep learning, Feature recombination

摘要： 针对微表情亚像素级形变易被全局噪声淹没、标准Transformer在小样本下易发生注意力稀释且深层特征趋于同质化等问题，本文提出一种融合动态Token修剪与层级特征重组的分层Transformer微表情识别方法。在预处理阶段，引入运动放大与分区域余弦加权策略，以余弦相似度为门控系数自适应压低与表情无关的噪声区域权重。在特征编码阶段，设计三层异构Transformer架构，依次嵌入动态Token修剪模块以剪除低质量Token、温和压缩策略以建模跨区域肌肉联动，以及层级特征重组模块以打破深层同质化。在聚合阶段，采用卷积块注意力聚合层校准强化判别性运动响应。在MEGC 2019复合数据集上，UF1和UAR分别达到89.10%和88.73%；在CASME II、SAMM和SMIC子集上，UF1分别取得96.16%、85.26%和83.39%的最优结果，验证了所提架构的有效性。

关键词: 微表情识别, 光流分区, 分层Transformer, 深度学习, 特征重组