2026-10-06T18:15:20Z https://api.figshare.com/v2/oai

oai:figshare.com:article/33945082 2026-09-20T13:35:20Z category_28879 item_type_3 month_year_09_2026

ALHCTNet Tang WeiCheng (23581150) Computer vision cnn ), learning Micro-expression recognition plays an important role in fields such as interpersonal communication, emotion analysis, and psychological research. At present, the micro-expression recognition task has achieved certain results, but maintaining a stable and high recognition accuracy is still a performance bottleneck problem of this task. To address this problem, we propose the Hierarchical CNN-Transformer Network Based on Automatic Localization (ALHCTNet) framework. This framework leverages three-dimensional optical flow maps to capture information about micro-expression motion sequences, and designs an automatic locator to select local features of five key facial regions in each optical flow map. Considering that the importance of each facial region may vary, we utilize a CNN+Transformer hybrid model. The CNN layer dynamically weights different channels, focusing on those channels beneficial for improving task performance while ignoring unimportant ones. Meanwhile, the Transformer layer attends to the interaction between features of different facial regions to enhance the model’s feature processing capability. We conduct a series of experiments on publicly available micro-expression Data Sets, including SAMM, SMIC, and CASME II. The experimental results show that the proposed method achieves the SOTA effect of microexpression three-classification tasks. The recognition accuracy of the training set is 99.39%, and the recognition accuracy of the test set is 100% on the mixed Data Set. 2026-09-20T13:35:20Z Dataset Dataset 10.6084/m9.figshare.33945082.v1 https://figshare.com/articles/dataset/ALHCTNet/33945082 CC BY 4.0