A deep learning-enhanced in-situ surface topography measurement method based on the focus variation microscopy and the industrial camera for material extrusion-based additive manufacturing

Focus variation microscopy is a powerful tool but is limited in its applicability to in-situ states. A research gap exists in adapting focus variation microscopy with inexpensive, easy-to-operate cameras to enable rapid surface topography acquisition in online measurements. To address this, we propose a novel deep learning-enhanced framework, M2CNet, in which images captured by a conventional industrial camera are first aligned with microscopy images using feature-based image registration. These aligned images are then paired with high-precision point clouds using a multi-focus window sliding technique and finally mapped to 3D point clouds via convolutional neural networks. A case study involving the surface of PLA fabricated by FDM showed that the M2CNet-16 model achieved the best result, with an average surface roughness (Sq) error of 6.4%, a Pearson correlation of 83.5%, and a processing time of 2.61 s. These results indicate that M2CNet improves training and prediction efficiency while maintaining state-of-the-art performance. Findings validate the feasibility of using simple cameras for high-precision topography measurements in material extrusion-based additive manufacturing.