Deep learning-driven antenna deformation measurement from single defocused images

The structural deformation of large high-precision reflector antennas caused by factors such as gravity, temperature, and wind load seriously deteriorates the electrical performance of the antenna. Rapid and effective measurement of reflector deformation is extremely important for compensating the electrical performance of the antenna. Traditional deformation measurement methods have limitations such as low efficiency or difficulty in meeting high-precision monitoring requirements. This article proposes a reflection surface antenna deformation measurement method based on a single-frame defocused intensity image and deep learning, aiming to use a single defocused far-field intensity image to recover the aperture field phase and obtain reflection surface deformation information. This method is based on the phase recovery technique of the deep residual network Res-Net, which actively introduces axial random deformation of the reflector to defocus the antenna and enhance the sensitivity of far-field intensity to deformation. The Res-Net model constructed adopts an encoder decoder structure, integrating improved residual Inception blocks and multi branch parallel convolution strategies to effectively capture multi-scale features and achieve complex nonlinear mapping from intensity to phase. The simulation results show that this method can effectively recover phase changes caused by typical deformation modes such as random surface shape error, gravity collapse, and large-scale continuous deformation. Its average structural similarity index (SSIM) is 0.9883, and the average root mean square error (RMSE) is only 0.00175, which is significantly better than the traditional Gerchberg-Saxton algorithm, and the calculation speed is improved by about 4 orders of magnitude. In addition, noise robustness testing shows that the model has good adaptability to Poisson noise, and phase recovery performance in Gaussian noise environments can be effectively improved through specific noise training. The proposed method provides a new technological approach for high-precision and rapid measurement of the deformation of reflector antennas.