Abstract: Objective　With the rapid advancement of 3D sensing technologies, point cloud data have been widely used in applications such as remote sensing, autonomous driving, and robotic navigation. Unlike 2D images, point clouds preserve precise geometric structures, offering unique advantages for representing complex scenes. However, their inherent unorderedness, sparsity, and non-uniform density pose significant challenges to feature extraction and classification. Although deep learning has made progress, existing methods still suffer from several critical limitations: inadequate capture of local feature, neglect of the importance of inter-channel features, difficulty in balancing geometric detail preservation and computational efficiency, and high complexity that limits their deployment on resource-constrained devices. To address these issues, this paper proposes a novel point cloud classification network that integrates local feature extraction with a complementary attention mechanism, aiming to achieve high accuracy, strong robustness, and low model complexity simultaneously. Methods　The proposed network is built upon two core modules: Local Feature Extraction Module (LFEM) and Sharp- Gentle Complementary Attention Module (SGCAM), as well as a voting-based classification mechanism, all of which work synergistically to extract discriminative features from raw point clouds. LFEM first applies the K-nearest neighbor algorithm to aggregate neighborhood features for each point, forming local patches that encode relative spatial relationships. Then, it introduces an Efficient Channel Attention mechanism that uses fast one-dimensional convolution to capture local crosschannel interactions without dimensionality reduction, thereby adaptively assigning weights along the channel dimension. This design enhances local structure representation while maintaining a lightweight architecture suitable for real-time applications. SGCAM, based on graph signal processing theory, constructs an adjacency matrix based on point cloud geometry and decomposes point cloud features into two complementary components: the sharp-variation (high-frequency) components corresponding to edges, corners, and rapid geometric changes, and the gentle-variation (low-frequency) components corresponding to smooth surfaces and gradual transitions. Individual attention weights for each component are computed using learnable nonlinear mappings implemented by multi-layer perceptrons, and then these weighted components are fused with the original point cloud features. This decomposition and reweighting strategy effectively reduces data redundancy, enhances the representation of critical geometric details, and captures complementary semantic information often missed by conventional attention mechanisms. The voting mechanism adopts a soft voting strategy under multiple random perturbations, including random scaling, translation, rotation, and additive noise, to average the output probability distributions of different perturbed versions of the same input, so as to improve classification stability and robustness, while reducing the impact of single prediction errors. Results and Discussion　Extensive experiments were conducted on two benchmark datasets: ModelNet10 (10 categories, 4, 899 models) and ModelNet40 (40 categories, 12, 311 models). The proposed method achieves overall classification accuracies of 96.1% on ModelNet10 and 93.0% on ModelNet40, outperforming mainstream models such as PointNet, PointNet++ , DGCNN, PointCNN, KPConv, and even the latest GateNet on ModelNet10. It is worth noting that although GateNet achieves a slightly higher accuracy (93.3%) on ModelNet40, its model complexity is significantly higher. Compared to GateNet, our method is more suitable for resource-constrained scenarios. Robustness tests demonstrate that when the number of input points is reduced from 1024 to 256, the classification accuracy of our method only drops by 1.6% (from 93.0% to 91.4%), while comparative methods such as Point-TnT and DGCNN decline by 5.9% and 45.8% respectively under the same conditions, indicating that our method has strong resistance to sparsity and noise. Model complexity analysis reveals that the proposed network only has 0.94 million (M) parameters and 20.30 GFLOPs of floating-point operations, achieving an excellent trade-off between accuracy and computational cost compared to DGCNN (1.81M parameters) and Point-TnT (3.94M parameters). The training time of our method on ModelNet40 is only 11 hours, which is notably shorter than that of PointNet++ (20 hours), PointCNN (15 hours), and DGCNN (25 hours), demonstrating efficient convergence and fast training of our method. Ablation studies confirm that LFEM, SGCAM, and the voting mechanism all contribute significantly to overall performance, and their synergistic integration forms an efficient system for feature extraction and decision optimization. Specifically, removing LFEM reduces accuracy by 1.5%, removing SGCAM reduces accuracy by 2.0%, and removing the voting mechanism reduces accuracy by 0.4%. Conclusion　The network proposed in this work effectively addresses key limitations of existing methods, including inadequate local feature capture, neglect of inter-channel importance, difficulty in preserving geometric structures, and high complexity. Through the synergistic integration of Local Feature Extraction Module (LFEM), Sharp-Gentle Complementary Attention Module (SGCAM), and soft voting, the method achieves a favorable balance between accuracy, robustness, and complexity. Experimental results on ModelNet10 and ModelNet40 demonstrate its superiority over many existing approaches. This work provides an efficient and accurate solution for 3D point cloud classification, with broad applications in autonomous driving, remote sensing, and robotic perception. Future work will explore more efficient modules, learning strategies, and model compression techniques, including knowledge distillation, pruning, and quantization, to reduce deployment requirements on edge devices.

Key words: laser radar, point cloud classification, local feature extraction, complementary attention 708