Due to the subtle and localized nature of micro-expressions, existing recognition methods often struggle to achieve high accuracy and efficiency. To address this issue, this paper introduces a novel recognition approach that minimizes the influence of static regions and enhances motion-relevant features. The method employs differential techniques to identify dynamic regions in the image, constructs a dynamic region mask, and enhances motion within the masked area. It suppresses amplification in static regions and reduces noise generation. The enhanced image serves as input to an optimized version of the CrossViT model tailored for micro-expression recognition. To evaluate the effectiveness of the proposed method, experiments were conducted on three widely used benchmark datasets: SMIC, CASME, and CASME II. The results show that the proposed approach achieves a performance gain of around 7.25% in accuracy compared to the widely adopted MERASTC method. This indicates that the algorithm has superior facial feature extraction capabilities for micro-expressions and achieves higher recognition accuracy.