TDS-Net: A High-Precision and Efficient Task-Driven Sampling Network for Freeform Surfaces Measurement

Coordinate measuring machines (CMMs) are widely used in industrial freeform surface measurement due to their high precision. The measurement process primarily involves point sampling and surface reconstruction. In this process, the choice of sampling method directly affects measurement accuracy and efficiency. However, existing sampling methods lack task-specific optimization for downstream surface reconstruction, making it difficult to achieve optimal sampling. To address these challenges, this paper proposes a task-driven adaptive sampling network (TDS-Net), which optimizes sampling for surface reconstruction in an end-to-end manner. TDS-Net consists of three core modules: the Multi-Channel Fusion Module (MCFM), the Adaptive Sampling Module (ASM), and the Downstream Task Module (DTM). First, the freeform surface point cloud is mapped to a 2D height map and combined with Gaussian curvature information, which is fed into MCFM to fuse multi-scale geometric features and identify key regions. ASM then employs an encoder-decoder architecture to extract deep feature information and generate a sampling probability map. A reparameterization technique decouples the discrete sampling from the computational graph, ensuring gradient propagation and training stability. Finally, DTM performs surface reconstruction based on sampled points and jointly optimizes task loss and uniformity loss, improving reconstruction accuracy while preventing local clustering of sampling points. The experimental results show that, for freeform surface reconstruction, TDS-Net reduces RMSE and PV by 26.75% and 36.03%, respectively, compared with existing sampling methods. It achieves high reconstruction accuracy while maintaining robust and adaptive sampling strategies, offering an efficient and intelligent solution for complex surface measurement tasks.