As an important cultural heritage of ancient China, oracle bone inscriptions have high historical and academic value. The unavoidable noise pollution during the topography process affects the subsequent research and analysis. We propose a method of oracle bone topography based on dynamic convolution and dimensional joint attention. Firstly, our proposed model connects two network branches in parallel. Secondly, the upper branch consists of multiple convolutional blocks to extract feature information at different scales. Then, the lower branch consists of multi-dimensional dynamic convolution and dimensional joint attention for capturing local details as well as global patterns. Finally, we introduce a context broadcast feedforward network between the upper and lower branches. It captures the contextual information of the oracle images and improves the learning ability of complex oracle features and noise patterns. The experimental data show that our method improves 1.21dB and 0.009 in terms of PSNR and SSIM, respectively, compared to AirNet on the homemade oracle bone topography image dataset OBD at a noise level of 50.