In X-ray fluorescence computed tomography (XFCT) imaging, the absorption attenuation of incident X-rays and fluorescent X-rays by the sample is a critical factor that restricts high-quality image reconstruction. This study proposes a deep-learning-based self-absorption correction method for XFCT, which utilizes a convolutional neural network based on U-Net to learn the symmetric structure distribution in the original projection data and recover complete projection data from the sinograms affected by self-absorption. Through numerical simulation, a fan-beam XFCT imaging system was established to obtain 20000 sets of fluorescence sinograms, which were then used for network training, testing, and validation. The projection data affected by self-absorption were further validated through a simulation using Geant4 software. The results indicate that the well-trained neural network can achieve self-absorption correction on incomplete projection data, thereby improving the quality of reconstructed images.