A convolutional neural network-based deep learning technique is proposed for intelligent OSNR estimation on eye diagrams. Compared with other machine learning methods, the optimal performances and 100% accuracies are achieved for four widely-used modulation formats.