Tabulated data has been widely used to facilitate systematic and intuitive management. In particular, tabular images that contain a few simple symbols are useful for maintaining mechanical systems. Several companies have accumulated tabular images as their property. Although these images are valuable as they can be used to solve difficult problems using data-based methods, such as deep learning, they still remain unavailable because it is expensive to digitize them. For these reasons, we propose a model comprised of a convolutional neural network (CNN) and fully convolutional network (FCN) to digitize tabular images. We used some ResNet components as they are well-suited to the characteristics of tabular image data. A training set for each model was constructed by writing symbols in blank tables and then augmenting them. As a result, the trained CNN and FCN models exhibited 99.2% and 97.7% accuracy in 4.75s and 0.132s of inference time, respectively.