Prosformer: Accurate Surface Reconstruction for Sparse Profilometer Measurement with Transformer

Surface micro-structure measurement is significant for precision manufacturing. However, existing stylus profilometer is inefficient, and sparse line-scan measurement can't support accurate surface description. To improve the reconstruction performance, we propose a high-accurate reconstruction method with sparse line-scan measurement based on attention mechanism. We first arrange the sparse-line measurement in the 2D matrix and crop as patch region. Then we utilize transformer to construct semantic relationships between patches and assign new weights to each patch to accurately model the structural relationships of target region and perform feature extraction, where the self-attention can enhance the description of local details while cross-attention will interact with global information. Finally, a fully connected network as a decoder is adopted to reconstruct accurate surface details with complete geometric representation. We refer to this model as Prosformer. Furthermore, we simulate a larger-scale surface micro-structure dataset to drive the training process and measure micro-structures to valid Prosformer. Experiments show proposed method can effectively restore complex surface details.