Micro-expressions, as indicators of true emotions, have significant applications in medical care and public safety. These expressions are characterized by their short duration, low intensity, and localized occurrence. These characteristics lead to the small sample problem in micro-expressions, making feature learning challenging and limiting the improvement of recognition performance. To address this issue, we propose a multimodal contrastive learning pre-training model based on Action Unit (AU) intensity perception. We conducted an experiment to determine the minimum threshold for recognizing facial expressions. Using this threshold, we filtered a large volume of unsupervised samples. The first stage involves unsupervised multimodal contrastive learning, where the model learns from differences in facial actions across various modalities. Subsequently, the model is trained on the micro-expression recognition task using a small amount of labeled data, overcoming the limitations of small sample sizes. Comparative experiments using the MEGC2019-CD and the multimodal dataset CAS(ME)3 datasets demonstrate the superiority of our method.