Abstract: Microwave holographic measurement technology is widely utilized in the surface measurement of telescopes, among which the phase recovery method is extensively applied in the surface calibration of radio telescopes due to its elimination of the need for additional specialized equipment. This method typically employs inversion algorithms to iteratively obtain an approximate near-field phase from far-field intensity data of the antenna. To enhance computational efficiency, this paper adopts deep learning techniques and designs a model based on the Conditional Generative Adversarial Network (CGAN) approach to address the issue of near-field phase recovery for reflector antennas under single-input far-field amplitude conditions. The phase recovery method proposed by this model dispenses with the traditional reliance on prior knowledge and the time-consuming iterative process. The original CGAN loss function has been augmented by incorporating Mean Squared Error (MSE) and Structural Similarity Index (SSIM) loss functions to optimize network training and improve the precision of phase recovery. Validation has shown that the CGAN network is robust against Poisson noise and can serve as a denoising tool in the phase recovery process. The CGAN framework not only enhances the precision of phase recovery and reduces computational complexity but also contributes to solving phase recovery issues in Fourier imaging systems. Moreover, this method can be extended to phase error measurement in other fields.