Three-stage training strategy phase unwrapping method for high speckle noises

Abstract

Deep learning has been widely used in phase unwrapping. However, owing to the noise of the wrapped phase, errors in wrap count prediction and phase calculation can occur, making it challenging to achieve high measurement accuracy under high-noise conditions. To address this issue, a three-stage multi-task phase unwrapping method was proposed. The phase retrieval was divided into three training stages: wrapped phase denoising, wrap count prediction, and unwrapped phase error compensation. In the first stage, a noise preprocessing module was trained to reduce noise interference, thereby improving the accuracy of the wrap count prediction and phase calculation. The second stage involved training the wrap count prediction module. A residual compensation module was added to correct the errors from the denoising results generated in the first stage. Finally, in the third stage, the phase error compensation module was trained to correct errors in the unwrapped phase calculated in the second stage. Additionally, a convolution-based multi-scale spatial attention module was proposed, which effectively reduces the interference of spatially inconsistent noise and can be applied to a convolutional neural network. The principles of the multi-task phase unwrapping method based on a three-stage training strategy were first introduced. Subsequently, the framework and training strategies for each stage were presented. Finally, the method was tested using simulated data with varying noise levels. It was compared with TIE, iterative TIE, the least squares phase unwrapping method, UNet, phaseNet2.0, and DeepLabV3 + with a phase correction operation, demonstrating the noise robustness and phase retrieval accuracy of the proposed method.