Self-Learning based Highly Efficient Sampling Strategy for Complex Surface Reconstruction on Contact Measurements

Contact measurements are significant for surface metrology and can provide highly precise results. However, the point-by-point touch sampling process is less efficient, which seriously limits their applications in manufacture process, especially for the measurement of multi-scale complex workpieces. On the other hand, the lack of high-quality labeled datasets in manufacturing industries prevents advanced supervised learning approaches from modeling and accelerating the measurement process. To address these problems, this paper proposed a highly efficient sparse sampling strategy to accelerate the measurement efficiency and a self-learning based approach to reconstruct precise dense results, that can not only dramatically reduce the number of sampling points but also eliminate the dataset demand to train the reconstruction algorithm. The proposed method can learn the prior of sparse samples and then reconstruct dense accurate measurements with self-supervised behavior based on the optimization process of encoder-decoder convolutional neural networks. Intensive experiments show that the proposed approach outperforms blind interpolated methods and even close to supervised learning approaches.