In this work, we discuss how to train convolutional neural networks to classify seismic images. We present a methodology to process and organize post-stack data into appropriate data sets for training and testing the model. We generated a few data sets by varying several parameters and analyzed the effects of those modifications on the performance of the model. In our experiments, we simulated the workflow using real data where the expert feeds the system with some interpreted lines from a cube, and a CNN classifies the remaining lines. We used two public seismic data sets: the Netherlands Offshore in F3 block and Penobscot. Finally, we obtained up to 99\% of accuracy using less than 5\% of the available data for training. It is important to highlight that the model had a good performance in identifying the main portions of the seismic images and distinguishing the layer related to salt deposit in Netherlands.