Video streaming analytic pipelines are deployed widely nowadays, where in many cases videos are transmitted continuously from cameras to servers. However, video analytic accuracy highly depends on video resolution, and when bandwidth suffers only low-resolution videos can be transmitted which leads to low accuracy. The rapid development of super resolution DNNs sparks the possibility of reconstructing high-resolution videos reliably from low resolution videos, and thus boosts video analytic accuracy. However, pretrained SR models may fail to cope with dynamic video content, while introducing online learning can be efficient under this condition. In this paper, we propose a practical system, LENS, which integrates an online learning period into analytic pipelines. The experimental results show that LENS is capable of responding to video dynamics and retraining SR models for changes quickly. In summary, LENS can save bandwidth consumption by up to 87%, or achieve a higher accuracy by up to 9.6% compared with state-of-the-art methods with slight additional bandwidth.