Abstract: Micro-expression recognition （MER） aims to reveal the hidden， true emotions of targets and is therefore of great significance in fields such as human-computer interaction， psychology， and security. However， micro-expressions occurred with weak intensity， transience of duration， and long-range dependence between facial motion units， making it difficult for traditional convolutional neural networks to effectively represent the inherent dynamic features of micro-expressions. In response to these issues， this paper proposes a micro-expression recognition algorithm based on a visual transformer and motion feature selection. The proposed algorithm first computes horizontal and vertical optical flow motion maps to describe facial motion using the TV-L1 and then encodes the relationships between motion units using a visual transformer. Next， this study introduced a feature selection and fusion module （FSFM） to effectively capture the key local information of micro-expressions and integrated a spatial consistency attention module（SCAM） to ensure spatial distribution consistency among different motion patterns. Finally， a cross attention fusion module（CAFM） was introduced to enhance micro-expression semantic information. Extensive experiments were conducted on three benchmark datasets， namely， MMEW， CASMEII， and SAMM. The proposed method achieved recognition accuracy values of 67.8%， 73.3%， and 68.7%， respectively. Compared with existing methods， the proposed algorithm demonstrates a significant improvement in accuracy in micro-expression recognition tasks， thereby further validating the effectiveness and superiority of the method.