Abstract: Millimeter-wave radar near-field imaging and target recognition have significant application potential in the security field. However， traditional recognition methods based on convolutional neural networks usually involve high computational complexity， which limits their practical deployment on terminal devices. To address this issue， this study proposes a lightweight target recognition framework that integrates sparse representation with point-cloud processing. First， the radar images are sparsely reconstructed using the subsampled random Fourier transform-orthogonal matching pursuit algorithm. Salient key points are extracted to construct two-dimensional （2D） sparse point clouds， thereby reducing the input size and data redundancy while preserving the key target features. Second， an efficient classification network named PointSENet++ is designed， considering the representation characteristics of sparse and non-uniform 2D point clouds. The network can effectively fuse local geometric structure information with global salient features by introducing a multi-scale feature extraction module and channel attention mechanism. Finally， experiments are conducted on a near-field imaging dataset constructed using the TI AWR1642BOOST millimeter-wave radar and a three-axis sliding platform. The results demonstrate that the proposed method achieves a classification accuracy of 98.84% with only 1.79M parameters， and its overall performance surpasses that of various conventional image- and point-cloud-based baseline models. The experimental results verify that the proposed method achieves high recognition accuracy on the self-constructed dataset， while also exhibiting favorable lightweight advantages in terms of parameter scale and computational cost， demonstrating its potential for embedded deployment in practical security systems.