Deep-learning-based phase measuring deflectometry for rapid measurement of specular components

High-precision three-dimensional measurement of specular components plays a critical role in precision manufacturing and defect detection. However, due to their high reflectivity and the time-consuming nature of data acquisition and processing, achieving fast and real-time measurement remains a challenge. To address this issue, we propose a phase measuring deflectometry method based on a deep learning framework, which enables single-frame phase unwrapping. Specifically, a cross-attention skip connection is introduced into the TransUNet architecture to selectively fuse encoder and decoder features. In terms of the loss function, a combination of binary cross-entropy, Dice loss, and deep supervision is adopted. The model ultimately regresses the wrapped phase components M and D. Both qualitative and quantitative experiments validate that the proposed method not only improves performance metrics and phase reconstruction accuracy but also significantly reduces error while maintaining real-time inference efficiency. Moreover, the method demonstrates strong applicability not only to highly reflective specular objects but also to low-reflectivity targets for phase unwrapping inference.