Deep Convolutional Neural Networks (DCNNs) correspond to the state-of-art for image classification. However to train such systems it is necessary to have access to a large number of samples and powerful computational resources, given the huge number of involved parameters. In the field of seismic images, large and freely available databases are scarce due to their strategic interest. In this situation, large architectures lead to hardly tractable problems in terms of overfitting. In this paper, we propose a reduced-size CNN with low computational cost that allows high accuracy performance on two small seismic datasets. The results are compared with KNN, SVM and LeNet.