Optical measurement techniques, such as interferometry, moiré techniques, and digital holography, are the most popular noncontact approaches for measuring three-dimensional (3D) object surfaces in terms of non-invasive, fast, and accurate evaluation. Usually, the property of the measured quantity is encoded in the phase of the intensity distribution of the fringe pattern, which can be decoded by phase retrieval, in other words, the recovery of a complex-valued signal from the sampled intensity patterns. In this way, phase demodulation of the fringe pattern plays a crucial role in the ubiquitous optical measurements. Among various single frame phase demodulation techniques, the high-frequency fringe pattern demodulation technique, such as Fourier transform profilometry, sampling moiré method and spatial carrier phase-shifting have been intensively studied and are mainly based on known analytical models of measurement systems, such as harmonic representation of the intensity of fringe patterns. But for low-frequency fringe pattern, phase reconstruction from only a single interferogram is difficult, especially for those including closed fringes. Sign ambiguity during the single-frame demodulation is one of the main problems that impede the development of single-frame interferometry. In this case, fringe pattern orthogonalization plays a very important role in low-frequency fringe pattern phase extraction. However, due to the ill-posed problem of orthogonalization of a single frame fringe pattern, the development of an analytical method for fringe pattern orthogonalizing is full of challenges. In recent years, researchers have demonstrated that deep learning is a powerful machine learning technique that uses artificial neural networks with deep layers to fit complex mathematical functions, thereby, deep learning provides a promising improvement over classical methods derived from explicit analytical formulations of the forward models. More specifically, deep learning approaches handle problems by searching and establishing sophisticated mapping between the input and the target data owing to the powerful computation capability, and therefore may provide a new solution for the phase demodulation of low-frequency fringe pattern. Inspired by recent successful artificial intelligence-based optical imaging applications, in this paper, we propose to utilize the deep learning to solve this problem of under sampling. This paper shows the new phase retrieval method based on deep learning can effectively improve performance and enable new functionalities for fringe profilometry. In the proposed network, the Generative Adversarial Nets areused to generate digitally the phase shifting of original image by combining the prior knowledge of network and fringe pattern denoising normalization. After training on labeled image pairs, the proposed method successfully implemente the desired phase-shifting fringes pattern, which can be viewed as the orthogonal transformation of a fringe pattern. With this Orthogonal transformation network, the wrapped phase can be extracted easily if the sampled fringes pattern is normalized using a trained deep neural network. The validity of the proposed Orthogonal transformation network is demonstrated on both the simulated and experimentally obtained fringe patterns. We also perform a comparative analysis of the proposed and existing approaches. Herein, we conducted fringe pattern denoising-normalization by using a deep-learning-based method developed because of its high-quality reconstruction ability. Thereafter, we input the normalized FP into the proposed Hilbert transformation network to perform Hilbert transform. We demonstrated our approach on both an open and a closed fringe pattern. Indeed, owing to local phase-sign ambiguity, the processed results show that the unwrapped phase map cannot be reconstructed adequately from the existing D4-PS wrapped map, even for a plane. Further, the reference phase from the proposed method is compared with the phase obtained by the multiple-frame high precision phase shift algorithm. Experimental results show that the proposed Orthogonal transformation network can provide a simple and robust solution for optical phase extraction from a single fringe pattern with phase error distribution within 0.05 rad and, therefore, make it allow for paving a new way to measure object 3D profilometry in a transient situation.