The traditional photon tomography method cannot be suitable for the objects containing heavy metal shielding. So in this study, based on the visual characteristics of photon tomography, pixel prediction based on unsupervised learning is carried out for the tomographic projection, and an encoder based on the surrounding information of metal shielding is constructed to generate the structure of the sheilded area. At the same time, on the basis of the generative adversarial network (GAN), the encoder is used as the generative network, and the discriminator is built for the training of the encoder. The results show that the reconstruction method based on pixel prediction can reconstruct the photonic tomography information in the area covered by heavy metal shielding, and the reconstruction results can accurately reflect the internal details of real objects, which indicates that the reconstruction method based on pixel prediction can effectively reduce the influence of metal structure in the object on the tomography results.