The phase unwrapping algorithm for path correlation is sensitive to strong noise, so it cannot unwrap the phase accurately. Therefore, in recent years, a phase unwrapping method based on deep learning has been proposed. Optical phase is usually characterized by Zernike polynomials. However, the training data obtained by simulation has class imbalance problem, which will lead to poor segmentation effect and even failure to demodulate subsequent package phases. A DeepLabv3+ network structure is presented with optical wrapping phase data as the training sample and different loss functions as the training samples. The experimental results show that the improved loss function can improve the segmentation accuracy to a certain extent and provide the possibility for the subsequent high-precision phase unwrapping.