Enhanced single-frame interferometry via hybrid conv-transformer architecture for ultra-precise phase retrieval

Precise dynamic single-frame interferometry based on virtual phase shifting technique remains challenging due to the difficulty in satisfying the requirements for the quality and amount of fine-grained fringe's interferograms. Here we introduce a novel deep learning architecture, the Transformer Encoder-Convolution Decoder Phase Shift Network (TECD-PSNet), that achieves high-fidelity interferogram reconstruction. TECD-PSNet seamlessly integrates the strengths of transformer blocks in capturing global descriptions and convolution blocks in efficient feature extraction. A key process is the incorporation of a residual local negative feedback enhancement mechanism that adaptively amplifies losses in high-error regions to boost fine-grained detail sensitivity. This approach enables accurate phase retrieval for diverse pupil shapes, enhancing adaptability to various optical setups, while significantly reducing the amount of training data required. Experiments demonstrate a 22.9% improvement in PSNR for reconstructed interferograms and a 36.7% reduction in RMS error for retrieved phases compared to state-of-the-art methods.