Multi task deep learning phase unwrapping method based on semantic segmentation

Abstract Phase unwrapping is a key step to obtain continuous phase distribution in optical phase measurement. When the wrapped phase obtained from the interference pattern is unclear and noisy, estimating the unwrapped phase becomes more challenging. As deep learning advances in optical image processing, it will enhance processing efficiency and accuracy, bringing broader possibilities for various applications. This paper introduces an innovative phase unwrapping method based on multi-task learning, aiming to simultaneously enhancing denoised images and predicting wrap count. The proposed network, named ICER-Net, comprises an encoder and two decoders, transforming the input low-luminance, noisy wrapped phase into two intermediate outputs: enhanced wrapped phase and wrap count. Finally, these two intermediate results are fused to obtain the unwrapped phase. Experimental results demonstrate that ICER-Net not only enhances the accuracy of phase unwrapping, particularly when facing challenges of various noise levels and luminance sizes but also exhibits outstanding performance in actual collected speckle phase images. This indicates that ICER-Net holds significant superiority in addressing complex issues in optical image processing.