3D surface imaging using structured-light illumination and cross-domain learning (Conference Presentation)

Deep learning has become a powerful tool in optics and photonics, particularly in optical metrology. However, its generalizability in 3D imaging is limited by reliance on data from the same domain, requiring identical systems and environments for training and testing. This limitation necessitates extensive data collection and retraining when conditions change, leading to high costs. To address this, an adaptive cross-domain learning framework with a dynamic synthesis strategy is proposed for structured-light 3D imaging. This framework dynamically extracts structured-light features and constructs new networks for each new input data, offering enhanced flexibility and accuracy across varied systems and environments. It excels in both out-of-distribution scenarios by leveraging expert neural networks and in in-distribution scenarios by surpassing state-of-the-art methods in accuracy.