Compared to digital ECG, paper ECG representations are still common in cardiac diagnosis and care. Therefore, in this work, we used deep learning methods to digitize and classify ECG images, which is an important step in analyzing cardiac data. Firstly, we used the UNet model to convert ECG images containing various complex situations into clean binarized images, and then vertically scanned the binarized images to extract ECG signals. In the classification task, we unified the images size and then inputted them into the ResNet50 model for training. We obtained an SNR of 2.78 for the reconstruction task and an F-measure of 0.82 for the classification task on our own validation set. Finally, in the unofficial stage, we obtained an SNR of -0.59 in the reconstruction task and an F-measure of 0.65 in the classification task.