Image-based visual servoing is challenging as the robot needs to locate the object and learn to control the arm in the image plane, which often undergo significant interference such as ambient light, distractions, and background clutter. Recent studies shows that the control policy can be efficiently learned by the reinforcement learning. In this paper, we present a data-driven image-based closed-loop visual servoing method via reinforcement learning algorithm without any prior knowledge of the task object or the intrinsic camera parameters. We first locate the object with a convolutional neural network backbone feature extraction network. Moreover the robot can determine the relative motion of the camera and servo the camera to the desired pose. We demonstrate that the reinforcement learning based approach is capable of steering the camera with only a single template image of the task object.