The ink penetration phenomenon caused by factors such as ink penetration, improper storage environment, paper quality, writing pressure and camera light seriously affects the readability and preservation quality of Tibetan document images. Although this problem has received increasing attention in the field of document processing and digitization, there is currently no publicly available dataset for the ink penetration phenomenon. To this end, we constructed a dataset of Tibetan documents with uniform illumination, Tibetan-Mulitlighting, by scanning documents and taking photos. We also supplemented the dataset by performing low-lighting and non-uniform illumination processing on some Tibetan document data images using a contrast adjustment strategy driven by prior knowledge and a method of illumination migration. A Transformer Tibetan document image enhancement method integrating Top-k selection is proposed. Using an encoder-decoder architecture, the encoder directly operates on pixel patches with position information, and selects the most relevant pixels in the process to improve the focus on global key text information. The decoder reconstructs a clean image from the encoded patches. Experiments show that compared with previous methods such as U-net, GAN, and DocEntr, the proposed model has improved FM, pFM, PSNR, and has a significant effect on reducing the DRD index. It is of great significance for constructing Tibetan document dataset and subsequent recognition of Tibetan document images.