ATCM-Net: A deep learning method for phase unwrapping based on perception optimization and learning enhancement

In electronic speckle pattern interferometry (ESPI), retrieving absolute phase from wrapped phase images captured by phase-shifting is challenging, especially for high speckle, greatly-variable density, discontinuities, and uneven background cases. These conditions are too complex for the common phase continuity assumption, and existing methods have limitations in phase retrieval. To address these limitations, we propose an approach enhancing phase unwrapping by compositing phase reconstruction, wrapped phase filtering, and wrap-count segmentation in one network architecture and training framework. We design a two-tasks coupling network (attention and transformer combined M network, ATCM-Net) using residual blocks, convolutional block attention modules, contextual transformers, dense blocks, and skip connections. We propose a composite loss function with reconstruction, filtering, and segmentation losses for effective training. We incorporate phase unwrapping theory into training to utilize prior information. We construct a dataset for phase unwrapping with the phase-shifting method, using realistic, varied samples sufficient for training. ATCM-Net is trained successfully with the proposed loss function on this dataset. We obtain the phase from wrapped phase effectively, without parameter fine-tuning or processing steps. We test our method on simulated and experimental samples, comparing it with existing methods and ATCM-Net trained by reconstruction and filtering loss. Performance is evaluated quantitatively and qualitatively for speckle reduction, retrieval accuracy, structure protection, and texture preservation. The results demonstrate our method’s superiority in phase retrieval, speckle reduction, structure protection, texture preservation, generalization, and batch performance on complex cases. Additionally, ATCM-Net shows good generalization when applied to phase retrieval in ESPI measurements and other technologies.