Recent advances in deep learning are driving the digitalization revolution, with the potential to automate many descriptive tasks in Earth and planetary sciences. Acquisition, description, and interpretation of geologic core is fundamental for subsurface characterization. However, it is time consuming, prone to individual bias or human error, and requires specialized knowledge. To address these challenges, we test the application of convolutional neural networks (CNNs) using semantic segmentation architectures to automate the identification of common lithofacies from core images. Images from more than 1100 m (3600 ft) of core from deep-water gravity-flow deposits in the Gulf of Mexico (GOM) and North Sea (NS) are used to train the CNNs. Validation accuracies of approximately 80% are achieved. Testing accuracies of approximately 60% are achieved by training with only data from the GOM and testing in the NS, demonstrating the potential for using trained networks in new settings and basins with relatively similar deposits. Misclassifications are mostly related to lithofacies classes that appear differently in the testing and validation data set due to the varying geological settings, or classes that have gradational boundaries and are geologically complex to describe. Results suggest semantic segmentation networks have significant potential to aid in the characterization and standardization of geologic core descriptions with acceptable accuracies at a fraction of the time required for conventional description. However, optimal deployment of these techniques in new data sets needs to be accompanied by efficient tools to allow geoscientists to interactively evaluate and edit the network predictions where required.