Most articles in the literature on electrocardiogram (ECG) arrhythmia classification use 1D convolutional networks. These networks only exploit relationships between left and right neighbors. 2D convolutional networks, on the other hand, also exploit relationships between upstream and downstream neighbors, allowing more complex relationships to be established among the data. To take advantage of this feature of 2D networks, this study proposes a new method for arrythmia classification using a 2D convolutional network. As input to this network, the one-dimensional ECG signal is converted into a 2D intensity image. Another contribution of the paper is proposing a new data augmentation technique that uses sliding windows. Data augmentation values up to 11times greater are achieved. Three CNN architectures and 3 optimizer methods (RMSProp, SGDM and ADAM) were evaluated. The best results obtained for 13 arrhythmia classifications were a global accuracy of 95.80%, which was obtained with the SGDM optimizer and data augmentation.