Abstract: ‍ ‍In the task of gesture classification based on millimeter-wave radar， the application of deep learning techniques can significantly improve accuracy. However， deep learning models heavily rely on a large amount of data， and when training samples are scarce， overfitting issues are prone to occur. Gathering gesture data using millimeter-wave radar can be time-consuming and labor-intensive， and there is often a limited amount of data available due to significant variations in millimeter-wave radar parameters. To address the issue of limited data， this study proposes a data augmentation method called Range-Doppler Image AutoEncoder with Attention Module （RDI-AEAM）， which incorporates an attention module to enhance the representation of millimeter-wave radar gesture data in range-Doppler image （RDI）. This method is designed to overcome the challenges posed by the lack of semantic information in RDIs， the difficulty of annotation and distinctive features. In the RDI-AEAM， a self-encoder network with an attention module is constructed. Firstly， a self-encoder is used to extract features and compress the data， learning the distribution of the input data and extracting useful features. Secondly， the attention module focuses on learning channel and spatial dimension features to address the problem of indistinct features， allowing the model to concentrate on important features. During the training process， predefined labels are assigned to the original data. The quality of the generated data is measured using the mean squared error loss function. When the generated data meets a predefined threshold， it is associated with the predefined labels， eliminating the need for additional post-labeling. By selecting 100% of the training set for augmentation， the accuracy of the RDI-AEAM improves compared to using only the training set for training. The augmented data results in an improvement of 0.83% in accuracy on our self-built dataset， 0.39% on the deepSoli dataset， and 3.23% on the VR-HGR dataset， indicating enhanced gesture discrimination performance. Furthermore， we investigate the effect of using even less original data for augmentation， augmenting only 25% of the training set， which yields improvements of 1.92%， 2.62%， and 1.56% on the three datasets， respectively.