UMSPU: Universal multi-size phase unwrapping network via mutual self-distillation and adaptive boosting ensemble segmenter

Abstract

Phase unwrapping is a crucial step in various high-precision measurement techniques. Deep learning - based methods are widely studied due to their better noise resistance and speed. However, existing phase unwrapping networks are constrained by the receptive field range and sparse semantic information, making them unable to effectively process high-resolution images, which severely limits their application in practical scenarios. To address this issue, we propose a Mutual Self-Distillation (MSD) mechanism and an adaptive-boosting ensemble segmenter to construct a Universal Multi-Size Phase Unwrapping network (UMSPU). MSD realizes cross-layer supervised learning by optimizing the bidirectional Kullback–Leibler divergence of attention maps, ensuring the precise extraction of fine-grained semantic features across different resolutions. The adaptive boosting ensemble segmenter combines weak segmenters with different receptive fields into a strong segmenter, ensuring stable segmentation at different spatial frequencies. The proposed mechanisms help UMSPU break the resolution limitations of previous networks, increasing the applicable resolution range from 256 $\times$ 256 to 2048 $\times$ 2048 (a 64-fold increase). It also enables the network to achieve highly robust effects in cross-domain applications with a lightweight architecture, taking only 22.66 ms to process a high-resolution image. This will effectively help the deep learning-based phase unwrapping method advance from the scientific research level to the practical application level.