A lightweight convolutional neural network (CNN) is presented in this study to automatically indentify atrial fibrillation (AF) from single-lead ECG recording. In contrast to existing methods employing a deeper architecture or complex feature-engineered inputs, this work presents an attempt to employ a lightweight CNN to confront current drawbacks such as higher computational requirement and inadequate training dataset, by using representative rhythms features of AF rather than raw ECG signal or hand-crafted features without any electrophysiological considerations. The experimental results suggested that this method presents the following significant advantages: (1) higher performances for indentifying AF in terms of accuracy, sensitivity, and specificity that are 97.5%, 97.8%, and 97.2%, respectively; (2) It is capable of automatically extracting the shared features of AF episodes of different patients and would be much robust and reliable; (3) with the cardiac rhythm features as input dataset, rather than complex transforming and classifying the raw data, thus requiring a lower computational resource. In conclusion, this automated method could analyze large amounts of data in a short time while assuring a relative high accuracy, and thus would potentially serve to provide a comfortable single-lead monitoring for patients and a clinical useful tool for doctors.