EAS-PointNetGPD: enhanced adaptive point cloud processing network for robotic arm grasping

In recent years, point cloud processing has garnered widespread attention in the fields of computer vision and robotics. As an innovative deep learning architecture, PointNet has successfully addressed the issues of disorder and irregularity in point cloud data. However, PointNet primarily relies on global features, neglecting local structural information, and is sensitive to noise and outliers in the input data. In practical applications, point cloud data often exhibits uneven distribution, which affects PointNet's performance in handling complex shapes. To address these issues, this paper proposes an improved PointNet model, EASPointNet-GPD (EASPN-GPD), which introduces SimAM and external attention mechanisms to enhance the representation capability of point cloud features. Experimental results show that EASPN-GPD achieves an accuracy of 87.22% on the YCB dataset for classification tasks, demonstrating significant performance improvements.