<p>Micro-expressions have made it difficult to solve the problems of diversity and balance due to the small dataset, while the properties of local regions of micro-expressions are also one of the problems that cannot be ignored. Therefore, to address these two problems, a convolutional network micro-expression recognition model with triple diversity feature augmentation is proposed. Among them, for the micro-expression local region problem, we combine local and global to construct fusion features by dividing each handcrafted feature into a local and global feature flow network. For the diversity and balance problems of micro-expressions, we first introduce an implicit semantic data augmentation algorithm to construct a diversity augmentation loss function for the semantic direction information of the samples. Meanwhile, considering the diversity value of the original sample, the original diversity cross-entropy loss function is constructed, which can make up for the loss of original diversity due to semantic diversity augmentation and ensure the effectiveness of feature diversity information. Then, considering the effect of class balance on diversity augmentation, the class-weighted mean feature is proposed for the first time, and the balanced diversity class loss function is constructed, which can improve the problems caused by balance. Finally, the three loss functions are fused to construct a triple diversity composite loss function. The experiments on micro-expression datasets and composite databases were conducted for 3- and 5-classification, respectively. The various experiments demonstrate the feasibility and effectiveness of our method while outperforming current state-of-the-art algorithms.</p>